Machine Learning For Finance
Machine Learning For Finance
Jannes Klaas
BIRMINGHAM - MUMBAI
Machine Learning for Finance
All rights reserved. No part of this book may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, without the prior written
permission of the publisher, except in the case of brief quotations embedded in
critical articles or reviews.
Every effort has been made in the preparation of this book to ensure the
accuracy of the information presented. However, the information contained in
this book is sold without warranty, either express or implied. Neither the author,
nor Packt Publishing or its dealers and distributors, will be held liable for any
damages caused or alleged to have been caused directly or indirectly by this book.
Packt Publishing has endeavored to provide trademark information about all of the
companies and products mentioned in this book by the appropriate use of capitals.
However, Packt Publishing cannot guarantee the accuracy of this information.
ISBN 978-1-78913-636-4
www.packtpub.com
mapt.io
Mapt is an online digital library that gives you full access to over 5,000 books
and videos, as well as industry leading tools to help you plan your personal
development and advance your career. For more information, please visit our
website.
Why subscribe?
• Spend less time learning and more time coding with practical eBooks
and Videos from over 4,000 industry professionals
• Learn better with Skill Plans built especially for you
• Get a free eBook or video every month
• Mapt is fully searchable
• Copy and paste, print, and bookmark content
Packt.com
Did you know that Packt offers eBook versions of every book published, with
PDF and ePub files available? You can upgrade to the eBook version at www.Packt.
com and as a print book customer, you are entitled to a discount on the eBook copy.
Get in touch with us at [email protected] for more details.
At www.Packt.com, you can also read a collection of free technical articles, sign up
for a range of free newsletters, and receive exclusive discounts and offers on Packt
books and eBooks.
Contributors
[i]
Table of Contents
[ ii ]
Table of Contents
Padding 80
Input shape 81
Simplified Conv2D notation 81
ReLU activation 81
MaxPooling2D 83
Flatten 84
Dense 84
Training MNIST 85
The model 85
Loading the data 87
Compiling and training 88
More bells and whistles for our neural network 90
Momentum90
The Adam optimizer 91
Regularization92
L2 regularization 93
L1 regularization 93
Regularization in Keras 94
Dropout 95
Batchnorm 97
Working with big image datasets 99
Working with pretrained models 101
Modifying VGG-16 104
Random image augmentation 105
Augmentation with ImageDataGenerator 105
The modularity tradeoff 107
Computer vision beyond classification 108
Facial recognition 109
Bounding box prediction 111
Exercises113
Summary 113
Chapter 4: Understanding Time Series 115
Visualization and preparation in pandas 117
Aggregate global feature statistics 118
Examining the sample time series 122
Different kinds of stationarity 126
Why stationarity matters 127
Making a time series stationary 127
When to ignore stationarity issues 128
Fast Fourier transformations 129
Autocorrelation131
Establishing a training and testing regime 134
[ iii ]
Table of Contents
[v]
Table of Contents
[ vi ]
Table of Contents
[ vii ]
Table of Contents
Summary420
Farewell421
Further reading 422
General data analysis 422
Sound science in machine learning 422
General machine learning 422
General deep learning 423
Reinforcement learning 423
Bayesian machine learning 423
Other Books You May Enjoy 425
Index429
[ viii ]
Preface
Aided by the availability of vast amounts of data computing resources, machine
learning (ML) has made big strides. The financial industry, which at its heart is
an information processing enterprise, holds an enormous amount of opportunity
for the deployment of these new technologies.
• Data scientists who want to break into finance and would like to know
about the spectrum of possible applications and relevant problems
• Developers in any FinTech business or quantitative finance professionals
who look to upgrade their skill set and want to incorporate advanced ML
methods into their modeling process
• Students who would like to prepare themselves for the labor market
and learn some practical skills valued by employers
This book assumes you have some working knowledge in linear algebra, statistics,
probability theory, and calculus. However, you do not have to be an expert in any
of those topics.
To follow the code examples, you should be comfortable with Python and the most
common data science libraries, such as pandas, NumPy, and Matplotlib. The book's
example code is presented in Jupyter Notebooks.
[ ix ]
Preface
Chapter 2, Applying Machine Learning to Structured Data, will deal with data that
resides in a fixed field within, for example, a relational database. We will walk
through the process of model creation: from forming a heuristic, to building a simple
model on engineered features, to a fully learned solution. On the way, we will learn
about how to evaluate our models with scikit-learn, how to train tree-based methods
such as random forests, and how to use Keras to build a neural network for this task.
Chapter 4, Understanding Time Series, looks at the large number of tools devoted to the
analysis of temporally related data. In this chapter, we will first discuss the "greatest
hits" that industry professionals have been using to model time series and how to use
them efficiently with Python. We will then discover how modern ML algorithms can
find patterns in time series and how they are complemented by classic methods.
Chapter 5, Parsing Textual Data with Natural Language Processing, uses the spaCy
library and a large corpus of news to discuss how common tasks such as named
entity recognition and sentiment analysis can be performed quickly and efficiently.
We will then learn how we can use Keras to build our own custom language models.
The chapter introduces the Keras functional API, which allows us to build much
more complex models that can, for instance, translate between languages.
Chapter 6, Using Generative Models, explains how generative models generate new
data. This is useful when we either do not have enough data or want to analyze our
data by learning about how the model perceives it. In this chapter, we will learn
about (variational) autoencoders as well as generative adversarial models. We will
learn how to make sense of them using the t-SNE algorithm and how to use them
for unconventional purposes, such as catching credit card fraud. We will learn about
how we can supplement human labeling operations with ML to streamline data
collection and labeling. Finally, we will learn how to use active learning to collect
the most useful data and greatly reduce data needs.
[x]
Preface
Chapter 8, Privacy, Debugging, and Launching Your Products, addresses how there is a lot
that can go wrong when building and shipping complex models. We will discuss how
to debug and test your data, how to keep sensitive data private while training models
on it, how to prepare your data for training, and how to disentangle why your model
is making the predictions it makes. We will then look at how to automatically tune
your model's hyperparameters, how to use the learning rate to reduce overfitting, and
how to diagnose and avoid exploding and vanishing gradients. After that, the chapter
explains how to monitor and understand the right metrics in production. Finally, it
discusses how you can improve the speed of your models.
Chapter 9, Fighting Bias, discusses how ML models can learn unfair policies and even
break anti-discrimination laws. It highlights several approaches to improve model
fairness, including pivot learning and causal learning. It shows how to inspect
models and probe for bias. Finally, we discuss how unfairness can be a failure in the
complex system that your model is embedded in and give a checklist that can help
you reduce bias.
Chapter 10, Bayesian Inference and Probabilistic Programming, uses PyMC3 to discuss the
theory and practical advantages of probabilistic programming. We will implement
our own sampler, understand Bayes theorem numerically, and finally learn how we
can infer the distribution of volatility from stock prices.
Equally, knowledge of Python and some popular data science libraries such as
pandas and Matplotlib is assumed.
[ xi ]
Preface
We also have other code bundles from our rich catalog of books and videos available
at https://github.com/PacktPublishing/. Check them out!
Conventions used
There are a number of text conventions used throughout this book.
CodeInText: Indicates code words in the text, database table names, folder names,
filenames, file extensions, pathnames, dummy URLs, user input, and Twitter
handles. For example; "Mount the downloaded WebStorm-10*.dmg disk image
file as another disk in your system."
[ xii ]
Preface
Bold: Indicates a new term, an important word, or words that you see on the screen,
for example, in menus or dialog boxes, also appear in the text like this. For example:
"Select System info from the Administration panel."
Get in touch
Feedback from our readers is always welcome.
General feedback: If you have questions about any aspect of this book, mention
the book title in the subject of your message and email us at customercare@
packtpub.com.
Errata: Although we have taken every care to ensure the accuracy of our content,
mistakes do happen. If you have found a mistake in this book we would be grateful
if you would report this to us. Please visit, http://www.packt.com/submit-errata,
selecting your book, clicking on the Errata Submission Form link, and entering
the details.
Piracy: If you come across any illegal copies of our works in any form on the
Internet, we would be grateful if you would provide us with the location address or
website name. Please contact us at [email protected] with a link to the material.
[ xiii ]
Preface
If you are interested in becoming an author: If there is a topic that you have
expertise in and you are interested in either writing or contributing to a book,
please visit http://authors.packtpub.com.
Reviews
Please leave a review. Once you have read and used this book, why not leave
a review on the site that you purchased it from? Potential readers can then see and
use your unbiased opinion to make purchase decisions, we at Packt can understand
what you think about our products, and our authors can see your feedback on their
book. Thank you!
[ xiv ]
Neural Networks and
Gradient-Based Optimization
The financial services industry is fundamentally an information processing industry.
An investment fund processes information in order to evaluate investments,
an insurance company processes information to price their insurances, while
a retail bank will process information in order to decide which products to offer
to which customers. It is, therefore, no accident that the financial industry was
an early adopter of computers.
The first stock ticker was the printing telegraph, which was invented back in 1867.
The first mechanical adding machine, which was directly targeted at the finance
industry, was patented in 1885. Then in 1971, the automatic teller banking machine,
which allowed customers to withdraw cash using a plastic card, was patented. That
same year, the first electronic stock exchange, the NASDAQ, opened its doors, and
11 years later, in 1982, the first Bloomberg Terminal was installed. The reason for
the happy marriage between the finance sector and computers is that success in the
industry, especially in investing, is often tied to you having an information advantage.
In the early days of Wall Street, the legends of the gilded age made brazen use
of private information. Jay Gould, for example, one of the richest men of his time,
placed a mole inside the US government. The mole was to give notice of government
gold sales and through that, tried to influence President Ulysses S. Grant as well
as his secretary. Toward the end of the 1930s, the SEC and CFTC stood between
investors and such information advantages.
[1]
Neural Networks and Gradient-Based Optimization
Most bank customers are everyday people as well, and increasingly, the main way
people are interacting with their bank, insurer, or pension is through an app on their
mobile phone.
In the decades before today, retail banks relied on the fact that people would
have to come into the branch, face-to-face, in order to withdraw cash or to make
a transaction. While they were in the branch, their advisor could also sell them
another product, such as a mortgage or insurance. Today's customers still want
to buy mortgages and insurance, but they no longer have to do it in person at
the branch. In today's world, banks tend to advise their clients online, whether
it's through the app or their website.
This online aspect only works if the bank can understand its customers' needs
from their data and provide tailor-made experiences online. Equally, from the
customers, perspective, they now expect to be able to submit insurance claims
from their phone and to get an instant response. In today's world, insurers need
to be able to automatically assess claims and make decisions in order to fulfill their
customers' demands.
This book is not about how to write trading algorithms in order to make a quick
buck. It is about leveraging the art and craft of building machine learning-driven
systems that are useful in the financial industry.
Building anything of value requires a lot of time and effort. Right now, the market
for building valuable things, to make an analogy to economics, is highly inefficient.
Applications of machine learning will transform the industry over the next few decades,
and this book will provide you with a toolbox that allows you to be part of the change.
Many of the examples in this book use data outside the realm of "financial data."
Stock market data is used at no time in this book, and this decision was made
for three specific reasons.
Firstly, the examples that are shown demonstrate techniques that can usually easily
be applied to other datasets. Therefore, datasets were chosen that demonstrate some
common challenges that professionals, like yourselves, will face while also remaining
computationally tractable.
Secondly, financial data is fundamentally time dependent. To make this book useful
over a longer span of time, and to ensure that as machine learning becomes more
prominent, this book remains a vital part of your toolkit, we have used some non-
financial data so that the data discussed here will still be relevant.
Finally, using alternative and non-classical data aims to inspire you to think about
what other data you could use in your processes. Could you use drone footage of
plants to augment your grain price models? Could you use web browsing behavior
to offer different financial products? Thinking outside of the box is a necessary skill
to have if you want to make use of the data that is around you.
[3]
Neural Networks and Gradient-Based Optimization
Supervised learning
Let's go back to our dog classifier. There are in fact many such classifiers currently
in use today. If you use Google images, for example, and search for "dog," it will use
an image classifier to show you pictures of dogs. These classifiers are trained under
a paradigm known as supervised learning.
[4]
Chapter 1
Supervised learning
If we have a high number of these labeled training examples, we can train a classifier
on detecting the subtle statistical patterns that differentiate dogs from all other animals.
Note: The classifier does not know what a dog fundamentally is. It only
knows the statistical patterns that linked images to dogs in training.
Unsupervised learning
While supervised learning has made great advances over the last few years, most
of this book will focus on working with labeled examples. However, sometimes we
may not have labels. In this case, we can still use machine learning to find hidden
patterns in data.
[5]
Neural Networks and Gradient-Based Optimization
Imagine a company that has a number of customers for its products. These customers
can probably be grouped into different market segments, but what we don't know is
what the different market segments are. We also cannot ask customers which market
segment they belong to because they probably don't know. Which market segment
of the shampoo market are you? Do you even know how shampoo firms segment
their customers?
This area of machine learning is far less developed than supervised learning,
but it still holds great potential.
Reinforcement learning
In reinforcement learning, we train agents who take actions in an environment,
such as a self-driving car on the road. While we do not have labels, that is, we cannot
tell what the correct action is in any situation, we can assign rewards or punishments.
For example, we could reward keeping a proper distance from the car in front.
Reinforcement learning
A driving instructor does not tell the student to "push the brake halfway down while
moving the steering wheel two degrees to the right," but rather they tell the student
whether they are doing well or not, while the student figures out the exact amount of
brakes to use.
Reinforcement learning has also made some remarkable progress in the past
couple of years and is considered by many to be a promising avenue toward
general artificial intelligence, that being computers that are as smart as humans.
[6]
Chapter 1
These bigger artificial neural networks were so effective that they got a name: deep
neural networks, or deep learning. Deep neural networks are especially good at
pattern detection. They can find complex patterns, such as the statistical pattern
of light and dark that describes a face in a picture, and they can do so automatically
given enough data.
This approach is superior if there is a very large number of rules, or even if these
rules are difficult to describe. Modern machine learning is, therefore, the ideal tool for
combing through the huge amounts of data the financial industry is confronted with.
There will come a time when even the most sophisticated deep neural network will
make a fundamentally wrong prediction, just as the advanced Collateralized Debt
Obligation (CDO) models did in the financial crises of 2008. Even worse, black box
machine learning models, which will make millions of decisions on loan approval or
insurance, impacting everyday people's lives, will eventually make wrong decisions.
[7]
Neural Networks and Gradient-Based Optimization
Sometimes they will be biased. Machine learning is ever only as good as the
data that we feed it, data that can often be biased in what it's showing, something
we'll consider later on in this chapter. This is something we must pay a lot of time
in addressing, as if we mindlessly deploy these algorithms, we will automate
discrimination too, which has the possibility of causing another financial crisis.
This is especially true in the financial industry, where algorithms can often have
a severe impact on people's lives while at the same time being kept secret. The
unquestionable, secret black boxes that gain their acceptance through the heavy
use of math pose a much bigger threat to society than the self-aware artificial
intelligence taking over the world that you see in movies.
While this is not an ethics book, it makes sense for any practitioner of the field
to get familiar with the ethical implications of his or her work. In addition to
recommending that you read Cathy O'Neil's Weapons of math destruction, it's also
worth asking you to swear The Modelers Hippocratic Oath. The oath was developed
by Emanuel Derman and Paul Wilmott, two quantitative finance researchers,
in 2008 in the wake of the financial crisis:
"I will remember that I didn't make the world, and it doesn't satisfy my equations.
Though I will use models boldly to estimate value, I will not be overly impressed
by mathematics. I will never sacrifice reality for elegance without explaining why
I have done so. Nor will I give the people who use my model false comfort about its
accuracy. Instead, I will make explicit its assumptions and oversights. I understand
that my work may have enormous effects on society and the economy, many of them
beyond my comprehension."
In recent years, machine learning has made a number of great strides, with
researchers mastering tasks that were previously seen as unsolvable. From
identifying objects in images to transcribing voice and playing complex board
games like Go, modern machine learning has matched, and continues to match
and even beat, human performance at a dazzling range of tasks.
Interestingly, deep learning is the method behind all these advances. In fact,
the bulk of advances come from a subfield of deep learning called deep neural
networks. While many practitioners are familiar with standard econometric models,
such as regression, few are familiar with this new breed of modeling.
The bulk of this book is devoted to deep learning. This is because it is one of the
most promising techniques for machine learning and will give anyone mastering
it the ability to tackle tasks considered impossible before.
In this chapter, we will explore how and why neural networks work in order
to give you a fundamental understanding of the topic.
[8]
Chapter 1
You can run Jupyter notebooks either on your local machine, on a server in the
cloud, or on a website such as Kaggle.
Note: All code examples for this book can be found here: https://
github.com/PacktPublishing/Machine-Learning-for-
Finance and for chapter 1 refer the following link: https://www.
kaggle.com/jannesklaas/machine-learning-for-finance-
chapter-1-code.
Deep learning is computer intensive, and the data used in the examples throughout
this book are frequently over a gigabyte in size. It can be accelerated by the use of
Graphics Processing Units (GPUs), which were invented for rendering video and
games. If you have a GPU enabled computer, you can run the examples locally. If you
do not have such a machine, it is recommended to use a service such as Kaggle kernels.
If you have many, long-running training jobs, it might be worth considering building
a "deep learning" box, a desktop computer with a GPU. There are countless tutorials
for this online and a decent box can be assembled for as little as a few hundred
dollars all the way to $5,000.
The examples in this book can all be run on Kaggle for free, though. In fact, they
have been developed using this site.
[9]
Neural Networks and Gradient-Based Optimization
After you've created your account, you can find the Kernels page by clicking
on Kernels located in the main menu, as seen in the following screenshot:
In the preceding screenshot, you can see a number of kernels that other people have
both written and published. Kernels can be private, but publishing kernels is a good
way to show skills and share knowledge.
To start a new kernel, click New Kernel. In the dialog that follows, you want to select
Notebook:
[ 10 ]
Chapter 1
You will get to the kernel editor, which looks like the preceding screenshot.
Note that Kaggle is actively iterating on the kernel design, and so a few elements
might be in different positions, but the basic functionality is the same. The most
important piece of a notebook is the code cells. Here you can enter the code and
run it by clicking the run button on the bottom left, or alternatively by pressing
Shift + Enter.
The variables you define in one cell become environment variables, so you can access
them in another cell. Markdown cells allow you to write text in markdown format to
add a description to what is going on in your code. You can upload and download
notebooks with the little cloud buttons featured in the top-right corner.
To publish a notebook from the kernel editor, firstly you must click the Commit
& Run button and then set the notebook to Public in the settings. To enable a GPU
on your notebook, make sure to check the Enable GPU button located in the bottom
right. It's important to remember that this will restart your notebook, so your
environment variables will be lost.
Once you run the code, the run button turns into a stop button. If your code ever
gets stuck, you can interrupt it by clicking that stop button. If you want to wipe
all environment variables and begin anew, simply click the restart button located
in the bottom-right corner.
With this system, you can connect a kernel to any dataset hosted on Kaggle, or
alternatively you can just upload a new dataset on the fly. The notebooks belonging
to this book already come with the data connection.
[ 11 ]
Neural Networks and Gradient-Based Optimization
Kaggle kernels come with the most frequently used packages preinstalled, so for
most of the time you do not have to worry about installing packages.
Sometimes this book does use custom packages not installed in Kaggle by default.
In this case, you can add custom packages at the bottom of the Settings menu.
Instructions for installing custom packages will be provided when they are used
in this book.
Kaggle kernels are free to use and can save you a lot of time and money, so it's
recommended to run the code samples on Kaggle. To copy a notebook, go to the
link provided at the beginning of the code section of each chapter and then click
Fork Notebook. Note that Kaggle kernels can run for up to six hours.
After installing Anaconda, you can start a Jupyter server locally by opening your
machine's Terminal and typing in the following code:
$ jupyter notebook
You can then visit the URL displayed in the Terminal. This will take you to your
local notebook server.
All code samples in this book use Python 3, so make sure you are using Python
3 in your local notebooks. If you are running your notebooks locally, you will also
need to install both TensorFlow and Keras, the two deep learning libraries used
throughout this book.
Installing TensorFlow
Before installing Keras, we need to first install TensorFlow. You can install
TensorFlow by opening a Terminal window and entering the following command:
$ sudo pip install TensorFlow
[ 12 ]
Chapter 1
For instructions on how to install TensorFlow with GPU support, simply click on this
link, where you will be provided with the instructions for doing so: https://www.
tensorflow.org/.
It's worth noting that you will need a CUDA-enabled GPU in order to run
TensorFlow with CUDA. For instructions on how to install CUDA, visit https://
docs.nvidia.com/cuda/index.html.
Installing Keras
After you have installed TensorFlow, you can install Keras in the same way,
by running the following command:
$ sudo pip install Keras
Keras will now automatically use the TensorFlow backend. Note that TensorFlow 1.7
will include Keras built in, which we'll cover this later on in this chapter.
Kaggle also offers a command-line interface, which allows you to download the
data more easily. Visit https://github.com/Kaggle/kaggle-api for instructions
on how to achieve this.
To run the notebooks for this book on an AMI, first set up the AMI, then download
the notebooks from GitHub, and then upload them to your AMI. You will have
to download the data from Kaggle as well. See the Using data locally section for
instructions.
[ 13 ]
Neural Networks and Gradient-Based Optimization
Approximating functions
There are many views on how best to think about neural networks, but perhaps the
most useful is to see them as function approximators. Functions in math relate some
input, x, to some output, y. We can write it as the following formula:
y = f ( x)
f ( x) = 4 ∗ x
In this case, we can give the function an input, x, and it would quadruple it:
y = f ( 2) = 8
You might have seen functions like this in school, but functions can do more;
as an example, they can map an element from a set (the collection of values the
function accepts) to another element of a set. These sets can be something other
than simple numbers.
imageContent = f (image)
This function would map an image of a cat to the label "cat," as we can see in the
following diagram:
[ 14 ]
Chapter 1
We should note that for a computer, images are matrices full of numbers and any
description of an image's content would also be stored as a matrix of numbers.
A neural network, if it is big enough, can approximate any function. It has been
mathematically proven that an indefinitely large network could approximate every
function. While we don't need to use an indefinitely large network, we are certainly
using very large networks.
Modern deep learning architectures can have tens or even hundreds of layers and
millions of parameters, so only storing the model already takes up a few gigabytes.
This means that a neural network, if it's big enough, could also approximate our
function, f, for mapping images to their content.
The condition that the neural network has to be "big enough" explains why deep
(big) neural networks have taken off. The fact that "big enough" neural networks
can approximate any function means that they are useful for a large number of tasks.
[ 15 ]
Neural Networks and Gradient-Based Optimization
A forward pass
Over the course of this book, we will build powerful neural networks that are able
to approximate extremely complex functions. We will be mapping text to named
entities, images to their content, and even news articles to their summaries. But for
now, we will work with a simple problem that can be solved with logistic regression,
a popular technique used in both economics and finance.
X1 X2 X3 y
0 1 0 0
1 0 0 1
1 1 1 1
0 1 1 0
y = X1
f ( X ) = X1
In this case, writing down the function is relatively straightforward. However, keep
in mind that in most cases it is not possible to write down the function, as functions
expressed by deep neural networks can become very complex.
For this simple function, a shallow neural network with only one layer will
be enough. Such shallow networks are also called logistic regressors.
[ 16 ]
Chapter 1
A logistic regressor
As we just explained, the simplest neural network is a logistic regressor. A logistic
regression takes in values of any range but only outputs values between zero
and one.
There is a wide range of applications where logistic regressors are suitable. One
such example is to predict the likelihood of a homeowner defaulting on a mortgage.
We might take all kinds of values into account when trying to predict the likelihood
of someone defaulting on their payment, such as the debtor's salary, whether they
have a car, the security of their job, and so on, but the likelihood will always be
a value between zero and one. Even the worst debtor ever cannot have a default
likelihood above 100%, and the best cannot go below 0%.
The following diagram shows a logistic regressor. X is our input vector; here
it's shown as three components, X1, X2, and X3.
W is a vector of three weights. You can imagine it as the thickness of each of the
three lines. W determines how much each of the values of X goes into the next
layer. b is the bias, and it can move the output of the layer up or down:
Logistic regressor
To compute the output of the regressor, we must first do a linear step. We compute
the dot product of the input, X, and the weights, W. This is the same as multiplying
each value of X with its weight and then taking the sum. To this number, we then
add the bias, b. Afterward, we do a nonlinear step.
[ 17 ]
Neural Networks and Gradient-Based Optimization
Since our dataset is quite small, we'll define it manually as NumPy matrices, as we
can see here:
X = np.array([[0,1,0],
[1,0,0],
[1,1,1],
[0,1,1]])
y = np.array([[0,1,1,0]]).T
[ 18 ]
Chapter 1
We can define the sigmoid, which squishes all the values into values between
0 and 1, through an activation function as a Python function:
def sigmoid(x):
return 1/(1+np.exp(-x))
So far, so good. We now need to initialize W. In this case, we actually know already
what values W should have. But we cannot know about other problems where we
do not know the function yet. So, we need to assign weights randomly.
The weights are usually assigned randomly with a mean of zero, and the bias
is usually set to zero by default. NumPy's random function expects to receive the
shape of the random matrix to be passed on as a tuple, so random((3,1)) creates
a 3x1 matrix. By default, the random values generated are between 0 and 1, with
a mean of 0.5 and a standard deviation of 0.5.
With that done, all the variables are set. We can now move on to do the linear step,
which is achieved with the following:
z = X.dot(W) + b
Now we can do the nonlinear step, which is run with the following:
A = sigmoid(z)
out:
[[ 0.60841366]
[ 0.45860596]
[ 0.3262757 ]
[ 0.36375058]]
But wait! This output looks nothing like our desired output, y, at all! Clearly,
our regressor is representing some function, but it's quite far away from the
function we want.
To better approximate our desired function, we have to tweak the weights, W, and
the bias, b. To this end, in the next section, we will optimize the model parameters.
[ 19 ]
Neural Networks and Gradient-Based Optimization
In other words, we need to look through the space of possible functions that can
be represented by our model in order to find a function, fˆ , that matches our desired
function, f, as closely as possible.
But how would we know how close we are? In fact, since we don't know f, we
cannot directly know how close our hypothesis, fˆ , is to f. But what we can do is
measure how well fˆ 's outputs match the output of f. The expected outputs of f
given X are the labels, y. So, we can try to approximate f by finding a function, fˆ ,
whose outputs are also y given X.
f (X )= y
fˆ ( X ) = yˆ
minimize
ˆ
D ( y, yˆ )
f ∈H
Within this formula, H is the space of functions that can be represented by our
model, also called the hypothesis space, while D is the distance function, which
we use to evaluate how close ŷ and y are.
Note: This approach makes a crucial assumption that our data, X, and
labels, y, represent our desired function, f. This is not always the case.
When our data contains systematic biases, we might gain a function
that fits our data well but is different from the one we wanted.
[ 20 ]
Chapter 1
As training data, you can use loan decisions made by human bank managers over
the years. However, this presents a problem as these managers might be biased. For
instance, minorities, such as black people, have historically had a harder time getting
a loan.
With that being said, if we used that training data, our function would also present
that bias. You'd end up with a function mirroring or even amplifying human biases,
rather than creating a function that is good at predicting who is a good debtor.
It is a commonly made mistake to believe that a neural network will find the intuitive
function that we are looking for. It'll actually find the function that best fits the data
with no regard for whether that is the desired function or not.
In this case, our issue is a binary classification problem. Because of that, we will
be using the binary cross-entropy loss, as we can see in the following formula:
1 N
DBCE ( y, yˆ ) = −
N
∑ y log ( yˆ ) + (1− y )log (1− yˆ )
i =1
i i i i
• DBCE: This is the distance function for binary cross entropy loss.
• : The loss over a batch of N examples is the average loss of all examples.
• yi ∗ log ( yˆi ) : This part of the loss only comes into play if the true value, y is 1. If
i
yi is 1, we want yˆl to be as close to 1 as possible, so we can achieve a low loss.
•
(1− yi ) log (1− yˆi ) : This part of the loss comes into play if y , is 0. If so, we want
i
yˆl to be close to 0 as well.
[ 21 ]
Neural Networks and Gradient-Based Optimization
The output, A, of our logistic regressor is equal to ŷ, so we can calculate the binary
cross-entropy loss as follows:
loss = bce_loss(y,A)
print(loss)
out:
0.82232258208779863
As we can see, this is quite a high loss, so we should now look at seeing how we can
improve our model. The goal here is to bring this loss to zero, or at least to get closer
to zero.
You can think of losses with respect to different function hypotheses as a surface,
sometimes also called the "loss surface." The loss surface is a lot like a mountain
range, as we have high points on the mountain tops and low points in valleys.
Our goal is to find the absolute lowest point in the mountain range: the deepest
valley, or the "global minimum." A global minimum is a point in the function
hypothesis space at which the loss is at the lowest point.
A "local minimum," by contrast, is the point at which the loss is lower than in the
immediately surrounding space. Local minima are problematic because while they
might seem like a good function to use at face value, there are much better functions
available. Keep this in mind as we now walk through gradient descent, a method for
finding a minimum in our function space.
Gradient descent
Now that we know what we judge our candidate models, fˆ , by, how do we tweak
the parameters to obtain better models? The most popular optimization algorithm
for neural networks is called gradient descent. Within this method, we slowly move
along the slope, the derivative, of the loss function.
Imagine you are in a mountain forest on a hike, and you're at a point where you've
lost the track and are now in the woods trying to find the bottom of the valley. The
problem here is that because there are so many trees, you cannot see the valley's
bottom, only the ground under your feet.
Now ask yourself this: how would you find your way down? One sensible approach
would be to follow the slope, and where the slope goes downwards, you go. This is
the same approach that is taken by a gradient descent algorithm.
To bring it back to our focus, in this forest situation the loss function is the
mountain, and to get to a low loss, the algorithm follows the slope, that is, the
derivative, of the loss function. When we walk down the mountain, we are updating
our location coordinates.
[ 22 ]
Chapter 1
The algorithm updates the parameters of the neural network, as we are seeing in the
following diagram:
Gradient descent
Gradient descent requires that the loss function has a derivative with respect to
the parameters that we want to optimize. This will work well for most supervised
learning problems, but things become more difficult when we want to tackle
problems for which there is no obvious derivative.
Gradient descent can also only optimize the parameters, weights, and biases of
our model. What it cannot do is optimize how many layers our model has or which
activation functions it should use, since there is no way to compute the gradient with
respect to model topology.
Backpropagation
Backpropagation allows us to apply gradient descent updates to the parameters of
a model. To update the parameters, we need to calculate the derivative of the loss
function with respect to the weights and biases.
[ 23 ]
Neural Networks and Gradient-Based Optimization
If you imagine the parameters of our models are like the geo-coordinates in our
mountain forest analogy, calculating the loss derivative with respect to a parameter
is like checking the mountain slope in the direction north to see whether you should
go north or south.
The following diagram shows the forward and backward pass through a logistic
regressor:
To keep things simple, we refer to the derivative of the loss function to any variable
as the d variable. For example, we'll write the derivative of the loss function with
respect to the weights as dW.
dy dy du
=
dx du dx
[ 24 ]
Chapter 1
The chain rule basically says that if you want to take the derivative through
a number of nested functions, you multiply the derivative of the inner function
with the derivative of the outer function.
This is useful because neural networks, and our logistic regressor, are nested
functions. The input goes through the linear step, a function of input, weights, and
biases; and the output of the linear step, z, goes through the activation function.
So, when we compute the loss derivative with respect to weights and biases, we'll
first compute the loss derivative with respect to the output of the linear step, z, and
use it to compute dW. Within the code, it looks like this:
dz = (A - y)
dW = 1/N * np.dot(X.T,dz)
db = 1/N * np.sum(dz,axis=0,keepdims=True)
Parameter updates
Now we have the gradients, how do we improve our model? Going back to our
mountain analogy, now that we know that the mountain goes up in the north
and east directions, where do we go? To the south and west, of course!
p = p − α ∗ dp
Here p is a model parameter (either in weight or a bias), dp is the loss derivative with
respect to p, and α is the learning rate.
The learning rate is something akin to the gas pedal within a car. It sets by how much
we want to apply the gradient updates. It is one of those hyperparameters that we
have to set manually, and something we will discuss in the next chapter.
[ 25 ]
Neural Networks and Gradient-Based Optimization
Firstly, we'll import numpy before we define the data. We can do this by running the
following code:
import numpy as np
np.random.seed(1)
X = np.array([[0,1,0],
[1,0,0],
[1,1,1],
[0,1,1]])
y = np.array([[0,1,1,0]]).T
The next step is for us to define the sigmoid activation function and loss function,
which we can do with the following code:
def sigmoid(x):
return 1/(1+np.exp(-x))
def bce_loss(y,y_hat):
N = y.shape[0]
loss = -1/N * np.sum((y*np.log(y_hat) + (1 - y)*np.log(1-y_hat)))
return loss
We'll then randomly initialize our model, which we can achieve with the
following code:
W = 2*np.random.random((3,1)) - 1
b = 0
As part of this process, we also need to set some hyperparameters. The first one
is alpha, which we will just set to 1 here. Alpha is best understood as the step size.
A large alpha means that while our model will train quickly, it might also overshoot
the target. A small alpha, in comparison, allows gradient descent to tread more
carefully and find small valleys it would otherwise shoot over.
The second one is the number of times we want to run the training process, also
called the number of epochs we want to run. We can set the parameters with the
following code:
alpha = 1
epochs = 20
[ 26 ]
Chapter 1
Since it is used in the training loop, it's also useful to define the number of samples in
our data. We'll also define an empty array in order to keep track of the model's losses
over time. To achieve this, we simply run the following:
N = y.shape[0]
losses = []
# Calculate loss
loss = bce_loss(y,A)
print('Epoch:',i,'Loss:',loss)
losses.append(loss)
# Calculate derivatives
dz = (A - y)
dW = 1/N * np.dot(X.T,dz)
db = 1/N * np.sum(dz,axis=0,keepdims=True)
# Parameter updates
W -= alpha * dW
b -= alpha * db
As a result of running the previous code, we would get the following output:
out:
Epoch: 0 Loss: 0.822322582088
Epoch: 1 Loss: 0.722897448125
Epoch: 2 Loss: 0.646837651208
Epoch: 3 Loss: 0.584116122241
Epoch: 4 Loss: 0.530908161024
Epoch: 5 Loss: 0.48523717872
Epoch: 6 Loss: 0.445747750118
Epoch: 7 Loss: 0.411391164148
Epoch: 8 Loss: 0.381326093762
Epoch: 9 Loss: 0.354869998127
Epoch: 10 Loss: 0.331466036109
Epoch: 11 Loss: 0.310657702141
[ 27 ]
Neural Networks and Gradient-Based Optimization
You can see that over the course of the output, the loss steadily decreases, starting
at 0.822322582088 and ending at 0.203407060401.
We can plot the loss to a graph in order to give us a better look at it. To do this,
we can simply run the following code:
import matplotlib.pyplot as plt
plt.plot(losses)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.show()
The output of the previous code, showing loss rate improving over time
[ 28 ]
Chapter 1
A deeper network
We established earlier in this chapter that in order to approximate more complex
functions, we need bigger and deeper networks. Creating a deeper network works
by stacking layers on top of each other.
In this section, we will build a two-layer neural network like the one seen in the
following diagram:
The input gets multiplied with the first set of weights, W1, producing an intermediate
product, z1. This is then run through an activation function, which will produce
the first layer's activations, A1.
These activations then get multiplied with the second layer of weights, W2,
producing an intermediate product, z2. This gets run through a second activation
function, which produces the output, A2, of our neural network:
z1 = X.dot(W1) + b1
a1 = np.tanh(z1)
z2 = a1.dot(W2) + b2
a2 = sigmoid(z2)
Note: The full code for this example can be found in the GitHub
repository belonging to this book.
[ 29 ]
Neural Networks and Gradient-Based Optimization
As you can see, the first activation function is not a sigmoid function but is actually
a tanh function. Tanh is a popular activation function for hidden layers and works
a lot like sigmoid, except that it squishes values in the range between -1 and 1 rather
than 0 and 1:
Backpropagation through our deeper network works by the chain rule, too. We go
back through the network and multiply the derivatives:
[ 30 ]
Chapter 1
Note that while the size of the inputs and outputs are determined by your problem,
you can freely choose the size of your hidden layer. The hidden layer is another
hyperparameter you can tweak. The bigger the hidden layer size, the more complex
the function you can approximate. However, the flip side of this is that the model
might overfit. That is, it may develop a complex function that fits the noise but not
the true relationship in the data.
Take a look at the following chart. What we see here is the two moons dataset that
could be clearly separated, but right now there is a lot of noise, which makes the
separation hard to see even for humans. You can find the full code for the two-
layer neural network as well as for the generation of these samples in the Chapter 1
GitHub repo:
[ 31 ]
Neural Networks and Gradient-Based Optimization
The following diagram shows a visualization of the decision boundary, that is, the
line at which the model separates the two classes, using a hidden layer size of 1:
As you can see, the network does not capture the true relationship of the data.
This is because it's too simplistic. In the following diagram, you will see the decision
boundary for a network with a hidden layer size of 500:
[ 32 ]
Chapter 1
This model clearly fits the noise, but not the moons. In this case, the right hidden
layer size is about 3.
Finding the right size and the right number of hidden layers is a crucial part of
designing effective learning models. Building models with NumPy is a bit clumsy
and can be very easy to get wrong. Luckily, there is a much faster and easier tool
for building neural networks, called Keras.
Importing Keras
When importing Keras, we usually just import the modules we will use. In this case,
we need two types of layers:
• The Dense layer is the plain layer that we have gotten to know in this chapter
• The Activation layer allows us to add an activation function
Keras offers two ways to build models, through the sequential and the functional
APIs. Because the sequential API is easier to use and allows more rapid model
building, we will be using it for most of the book. However, in later chapters,
we will take a look at the functional API as well.
[ 33 ]
Neural Networks and Gradient-Based Optimization
Stacking layers
Firstly, we create an empty sequential model with no layers:
model = Sequential()
Then we can add layers to this model, just like stacking a layer cake, with
model.add().
For the first layer, we have to specify the input dimensions of the layer. In our case,
the data has two features, the coordinates of the point. We can add a hidden layer of
size 3 with the following code:
model.add(Dense(3,input_dim=2))
Note how we nest the functions inside model.add(). We specify the Dense layer,
and the positional argument is the size of the layer. This Dense layer now only
does the linear step.
Then, we add the linear step and the activation function of the output layer in the
same way, by calling up:
model.add(Dense(1))
model.add(Activation('sigmoid'))
Then to get an overview of all the layers we now have in our model, we can use the
following command:
model.summary()
[ 34 ]
Chapter 1
Total params: 13
Trainable params: 13
Non-trainable params: 0
You can see the layers listed nicely, including their output shape and the number of
parameters the layer has. None, located within the output shape, means that the layer
has no fixed input size in that dimension and will accept whatever we feed it. In our
case, it means the layer will accept any number of samples.
In pretty much every network, you will see that the input dimension on the first
dimension is variable like this in order to accommodate the different amounts
of samples.
The simple optimizer we have used so far is called the Stochastic Gradient
Descent, or SGD. To look at more optimizers, see Chapter 2, Applying Machine
Learning to Structured Data.
The loss function we use for this binary classification problem is called binary
cross-entropy. We can also specify what metrics we want to track during training.
In our case, accuracy, or just acc to keep it short, would be interesting to track:
model.compile(optimizer='sgd',
loss='binary_crossentropy',
metrics=['acc'])
This will train the model for 900 iterations, which are also referred to as epochs.
The output should look similar to this:
Epoch 1/900
200/200 [==============================] - 0s 543us/step -
loss: 0.6840 - acc: 0.5900
Epoch 2/900
[ 35 ]
Neural Networks and Gradient-Based Optimization
Epoch 899/900
200/200 [==============================] - 0s 90us/step -
loss: 0.2900 - acc: 0.8800
Epoch 900/900
200/200 [==============================] - 0s 87us/step -
loss: 0.2901 - acc: 0.8800
The full output of the training process has been truncated in the middle, this is to
save space in the book, but you can see that the loss goes continuously down while
accuracy goes up. In other words, success!
Over the course of this book, we will be adding more bells and whistles to these
methods. But at this moment, we have a pretty solid understanding of the theory of
deep learning. We are just missing one building block: how does Keras actually work
under the hood? What is TensorFlow? And why does deep learning work faster on
a GPU?
We will be answering these questions in the next, and final, section of this chapter.
TensorFlow is a software library developed by Google and is very popular for deep
learning. In this book, we usually try to work with TensorFlow only through Keras,
since that is easier than working with TensorFlow directly. However, sometimes
we might want to write a bit of TensorFlow code in order to build more advanced
models.
The goal of TensorFlow is to run the computations needed for deep learning as
quickly as possible. It does so, as the name gives away, by working with tensors in a
data flow graph. Starting in version 1.7, Keras is now also a core part of TensorFlow.
[ 36 ]
Chapter 1
This book will treat Keras as a standalone library. However, you might want to use
a different backend for Keras one day, as it keeps the code cleaner if we have shorter
import statements.
The next tensor is a vector, also known as a rank-one tensor. The next The next ones
up the order are matrices, called rank-two tensors; cube matrices, called rank-three
tensors; and so on. You can see the rankings in the following table:
This book mostly uses the word tensor for rank-three or higher tensors.
TensorFlow and every other deep learning library perform calculations along
a computational graph. In a computational graph, operations, such as matrix
multiplication or an activation function, are nodes in a network. Tensors get
passed along the edges of the graph between the different operations.
A forward pass through our simple neural network has the following graph:
[ 37 ]
Neural Networks and Gradient-Based Optimization
The advantage of structuring computations as a graph is that it's easier to run nodes
in parallel. Through parallel computation, we do not need one very fast machine; we
can also achieve fast computation with many slow computers that split up the tasks.
This is the reason why GPUs are so useful for deep learning. GPUs have many small
cores, as opposed to CPUs, which only have a few fast cores. A modern CPU might
only have four cores, whereas a modern GPU can have hundreds or even thousands
of cores.
The entire graph of just a very simple model can look quite complex, but you can
see the components of the dense layer. There is a matrix multiplication (matmul),
adding bias and a ReLU activation function:
[ 38 ]
Chapter 1
Exercises
Now that we have finished the first chapter in this exciting journey, I've got
a challenge for you! You'll find some exercises that you can do that are all
themed around what we've covered in this chapter!
Summary
And that's it! We've learned how neural networks work. Throughout the rest of this
book, we'll look at how to build more complex neural networks that can approximate
more complex functions.
As it turns out, there are a few tweaks to make to the basic structure for it to work
well on specific tasks, such as image recognition. The basic ideas introduced in
this chapter, however, stay the same:
[ 39 ]
Neural Networks and Gradient-Based Optimization
The key takeaway from this chapter is that while we are looking for function f, we
can try and find it by optimizing a function to perform like f on a dataset. A subtle
but important distinction is that we do not know whether fˆ works like f at all. An
often-cited example is a military project that tried to use deep learning to spot tanks
within images. The model trained well on the dataset, but once the Pentagon wanted
to try out their new tank spotting device, it failed miserably.
In the tank example, it took the Pentagon a while to figure out that in the dataset they
used to develop the model, all the pictures of the tanks were taken on a cloudy day
and pictures without a tank where taken on a sunny day. Instead of learning to spot
tanks, the model had learned to spot grey skies instead.
This is just one example of how your model might work very differently to how you
think, or even plan for it to do. Flawed data might seriously throw your model off
track, sometimes without you even noticing. However, for every failure, there are
plenty of success stories in deep learning. It is one of the high-impact technologies
that will reshape the face of finance.
In the next chapter, we will get our hands dirty by jumping in and working with
a common type of data in finance, structured tabular data. More specifically, we will
tackle the problem of fraud, a problem that many financial institutions sadly have
to deal with and for which modern machine learning is a handy tool. We will learn
about preparing data and making predictions using Keras, scikit-learn, and XGBoost.
[ 40 ]
Applying Machine Learning
to Structured Data
Structured data is a term used for any data that resides in a fixed field within
a record or file, two such examples being relational databases and spreadsheets.
Usually, structured data is presented in a table in which each column presents
a type of value, and each row represents a new entry. Its structured format means
that this type of data lends itself to classical statistical analysis, which is also why
most data science and analysis work is done on structured data.
In day-to-day life, structured data is also the most common type of data available
to businesses, and most machine learning problems that need to be solved in
finance deal with structured data in one way or another. The fundamentals
of any modern company's day-to-day running is built around structured data,
including, transactions, order books, option prices, and suppliers, which are
all examples of information usually collected in spreadsheets or databases.
This chapter will walk you through a structured data problem involving credit card
fraud, where we will use feature engineering to identify the fraudulent transaction
from a dataset successfully. We'll also introduce the basics of an end-to-end (E2E)
approach so that we can solve common financial problems.
Fraud is an unfortunate reality that all financial institutions have to deal with. It's a
constant race between companies trying to protect their systems and fraudsters who
are trying to defeat the protection in place. For a long time, fraud detection has relied
on simple heuristics. For example, a large transaction made while you're in a country
you usually don't live in will likely result in that transaction being flagged.
[ 41 ]
Applying Machine Learning to Structured Data
Yet, as fraudsters continue to understand and circumvent the rules, credit card
providers are deploying increasingly sophisticated machine learning systems to
counter this.
In this chapter, we'll look at how a real bank might tackle the problem of fraud.
It's a real-world exploration of how a team of data scientists starts with a heuristic
baseline, then develops an understanding of its features, and from that, builds
increasingly sophisticated machine learning models that can detect fraud. While
the data we will use is synthetic, the process of development and tools that we'll
use to tackle fraud are similar to the tools and processes that are used every day
by international retail banks.
So where do you start? To put it in the words of one anonymous fraud detection
expert that I spoke to, "I keep thinking about how I would steal from my employer, and
then I create some features that would catch my heist. To catch a fraudster, think like a
fraudster." Yet, even the most ingenious feature engineers are not able to pick up
on all the subtle and sometimes counterintuitive signs of fraud, which is why the
industry is slowly shifting toward entirely E2E-trained systems. These systems, in
addition to machine learning, are both focuses of this chapter where we will explore
several commonly used approaches to flag fraud.
This chapter will act as an important baseline to Chapter 6, Using Generative Models,
where we will again be revisiting the credit card fraud problem for a full E2E model
using auto-encoders.
The data
The dataset we will work with is a synthetic dataset of transactions generated by
a payment simulator. The goal of this case study and the focus of this chapter is to
find fraudulent transactions within a dataset, a classic machine learning problem
many financial institutions deal with.
[ 42 ]
Chapter 2
The dataset we're using stems from the paper PaySim: A financial mobile money
simulator for fraud detection, by E. A. Lopez-Rojas, A. Elmir, and S. Axelsson. The
dataset can be found on Kaggle under this URL: https://www.kaggle.com/ntnu-
testimon/paysim1.
Before we break it down on the next page, let's take a minute to look at the dataset
that we'll be using in this chapter. Remember, you can download the data with the
preceding link.
step type amount nameOrig oldBalance newBalance nameDest oldBalance newBalance isFraud isFlagged
Orig Orig Dest Dest Fraud
As seen in the first row, the dataset has 11 columns. Let's explain what each one
represents before we move on:
[ 43 ]
Applying Machine Learning to Structured Data
In the preceding table, we can see 10 rows of data. It's worth noting that there
are about 6.3 million transactions in our total dataset, so what we've seen is a small
fraction of the total amount. As the fraud we're looking at only occurs in transactions
marked as either TRANSFER or CASH_OUT, all other transactions can be dropped,
leaving us with around 2.8 million examples to work with.
[ 44 ]
Chapter 2
Feature engineering-based models are a midway between data and humans shaping
rules, where human knowledge and creativity are exploited to craft good features,
and data and machine learning are used to create a model from those features.
E2E models learn purely from collected data without using expert knowledge.
As discussed before, this often yields much better results, but at the cost of taking
a lot of time to complete. This method also has some additional elements worth
considering. For instance, collecting the large amount of data that will be needed
is an expensive task, as humans have to label millions of records.
Though for many people in the industry right now, they take the view that shipping
a poor model is often better than not shipping anything at all. After all, having some
protection against fraud is better than simply having none.
Using a heuristic approach that lets through half of all fraudulent transactions is
better than having no fraud detection at all. The graph shows us the performance of
the three models we introduced earlier on, against the time taken to implement them.
The methods used and the performance of the system during development
The best method is to use a combination of all three. If we deploy a heuristic model
that meets the basic requirements of the task that it set out to achieve, then it can
be shipped. By employing this method, the heuristic then becomes the baseline that
any other approach has to beat. Once your heuristic model is deployed, then all your
efforts should then be directed toward building a feature-based model, which as
soon as it beats the initially deployed heuristic model, can then be deployed while
you continue to refine the model.
As we've discussed before, feature-based models often deliver pretty decent
performance on structured data tasks; this gives companies the time to undertake
the lengthy and expensive task of building an E2E model, which can be shipped
once it beats the feature-based model. Now that we understand the type of models
we're going to build, let's look at the software we need to build them.
[ 45 ]
Applying Machine Learning to Structured Data
The majority of this book is dedicated to working with the Keras library, while this
chapter makes extensive use of the other libraries mentioned. The goal here is less
about teaching you all the tips and tricks of all the different libraries, but more about
showing you how they are integrated into the process of creating a predictive model.
Note: All of the libraries needed for this chapter are installed on Kaggle
kernels by default. If you are running this code locally, please refer to
the setup instructions in Chapter 1, Neural Networks and Gradient-Based
Optimization, and install all of the libraries needed.
Let's start by defining a simple heuristic model and measuring how well it does at
measuring fraud rates.
[ 46 ]
Chapter 2
The following code will create a new column, Fraud_Heuristic, and in turn
assigns a value of 1 in rows where the type is TRANSFER, and the amount is more
than $200,000:
df['Fraud_Heuristic '] = np.where(((df['type'] == 'TRANSFER') &
(df['amount'] > 200000)),1,0)
With just two lines of code, it's easy to see how such a simple metric can be easy to
write, and quick to deploy.
The F1 score
One important thing we must consider is the need for a common metric on which
we can evaluate all of our models on. In Chapter 1, Neural Networks and Gradient-Based
Optimization, we used accuracy as our emulation tool. However, as we've seen, there
are far fewer fraudulent transactions than there are genuine ones. Therefore a model
that classifies all the transactions as genuine can have a very high level of accuracy.
One such metric that is designed to deal with such a skewed distribution is the F1 score,
which considers true and false positives and negatives, as you can see in this chart:
We can first compute the precision of our model, which specifies the share of predicted
positives that were positives, using the following formula:
TP
precision =
TP + FP
Recall measures the share of predicted positives over the actual number of positives,
as seen in this formula:
TP
recall =
TP + FN
[ 47 ]
Applying Machine Learning to Structured Data
The F1 score is then calculated from the harmonic mean, an average, of the two
measures, which can be seen in the following formula:
precision ∗ recall
F1 = 2 ∗
precision + recall
To compute this metric in Python, we can use the metrics module of scikit-learn,
or sklearn for short:
from sklearn.metrics import f1_score
Given the predictions we've made, we can now easily compute the F1 score using the
following command:
f1_score(y_pred=df['Fraud_Heuristic '],y_true=df['isFraud'])
out: 0.013131315551742895
Firstly, let's study the code appendix for the plot_confusion_matrix function:
from sklearn.metrics import confusion_matrix cm = confusion_matrix(
y_pred=df['Fraud_Heuristic '],y_true=df['isFraud'])
plot_confusion_matrix(cm,['Genuine','Fraud'])
[ 48 ]
Chapter 2
So, just how accurate was that model? As you can see in our confusion matrix, from
our dataset of 2,770,409 examples, 2,355,826 were correctly classified as genuine,
while 406,370 were falsely classified as fraud. In fact, only 2,740 examples were
correctly classified as fraud.
[ 49 ]
Applying Machine Learning to Structured Data
To put it frankly, our heuristic model has performed quite poorly, detecting only
34.2% of fraud, which is not good enough. So, using the other two methods in the
following sections, we're going to see whether we can do better.
As an example, in our fraud case, it seems intuitive that fraudsters will create new
accounts for their fraudulent schemes and won't be using the same bank account
that they pay for their groceries with.
Domain experts are able to use their extensive knowledge of a problem in order
to come up with other such examples of intuition. They'll know more about how
fraudsters behave and can craft features that indicate such behavior. All of these
intuitions are then usually confirmed by statistical analysis, something that can even
be used to open the possibilities of discovering new features.
Statistical analysis can sometimes turn up quirks that can be turned into predictive
features. However, with this method, engineers must beware of the data trap.
Predictive features found in the data might only exist in that data because any
dataset will spit out a predictive feature if it's wrangled with for long enough.
A data trap refers to engineers digging within the data for features forever, and
never questioning whether those features they are searching for are relevant.
Data scientists stuck in to the data trap keep euphorically finding features, only to
realize later that their model, with all those features, does not work well. Finding
strong predictive features in the training set is like a drug for data science teams.
Yes, there's an immediate reward, a quick win that feels like a validation of one's
skills. However, as with many drugs, the data trap can lead to an after-effect in
which teams find that weeks' or months' worth of work in finding those features
was actually, useless.
Take a minute to ask yourself, are you in that position? If you ever find yourself
applying analysis after analysis, transforming data in every possible way, chasing
correlation values, you might very well be stuck in a data trap.
[ 50 ]
Chapter 2
Humans can use their qualitative reasoning skills to avoid fitting noise, which is
a big advantage humans have over machines. If you're a data scientist, you should
use this skill to create more generalizable models.
The goal of this section was not to showcase all the features that feature engineering
could perform on this dataset, but just to highlight the three approaches and how
they can be turned into features.
The time steps in our dataset represent one hour. Therefore, we can generate the time
of the day by simply taking the remainder of a division by 24, as seen in this code:
df['hour'] = df['step'] % 24
From there, we can then count the number of fraudulent and genuine transactions
at different times. To calculate this, we must run the following code:
frauds = []
genuine = []
for i in range(24):
f = len(df[(df['hour'] == i) & (df['isFraud'] == 1)])
g = len(df[(df['hour'] == i) & (df['isFraud'] == 0)])
frauds.append(f)
genuine.append(g)
Then finally, we can plot the share of genuine and fraudulent transactions over the
course of the day into a chart. To do this, we must run the following code:
fig, ax = plt.subplots(figsize=(10,6))
ax.plot(genuine/np.sum(genuine), label='Genuine')
ax.plot(frauds/np.sum(frauds),dashes=[5, 2], label='Fraud')
plt.xticks(np.arange(24))
legend = ax.legend(loc='upper center', shadow=True)
[ 51 ]
Applying Machine Learning to Structured Data
The share of fraudulent and genuine transactions conducted throughout each hour of the day
As we can see in the preceding chart, there are much fewer genuine transactions at
night, while fraudulent behavior continues over the day. To be sure that night is a
time when we can hope to catch fraud, we can also plot the number of fraudulent
transactions as a share of all transactions. To do this, we must run the following
command:
fig, ax = plt.subplots(figsize=(10,6))
ax.plot(np.divide(frauds,np.add(genuine,frauds)), label='Share of
fraud')
plt.xticks(np.arange(24))
legend = ax.legend(loc='upper center', shadow=True)
The share of transactions that are fraudulent per hour of the day
[ 52 ]
Chapter 2
Once we run that code, we can see that at around 5 AM, over 60% of all
transactions seem to be fraudulent, which appears to make this a great time
of the day to catch fraud.
We can check whether there are fraudulent transfer destination accounts that are
the origin of the fraudulent cash outs by running the following code:
dfFraudTransfer = df[(df.isFraud == 1) & (df.type == 'TRANSFER')]
dfFraudCashOut = df[(df.isFraud == 1) & (df.type == 'CASH_OUT')]
dfFraudTransfer.nameDest.isin(dfFraudCashOut.nameOrig).any()
out: False
According to the output, there seems to be no fraudulent transfers that are the origin
of fraudulent cash outs. The behavior expected by the experts is not visible in our
data. This could mean two things: firstly, it could mean that the fraudsters behave
differently now, or secondly that our data does not capture their behavior. Either
way, we cannot use this insight for predictive modeling here.
To begin with, we can calculate the share of fraudulent transactions with this
property by running the following code:
dfOdd = df[(df.oldBalanceDest == 0) &
(df.newBalanceDest == 0) &
(df.amount)]
len(dfOdd[(df.isFraud == 1)]) / len(dfOdd)
out: 0.7046398891966759
As you can see, the share of fraudulent transactions stands at 70%, so this quirk
seems to be a good feature at detecting fraud in transactions. However, it is
important to ask ourselves how this quirk got into our data in the first place.
One possibility could be that the transactions never come through.
[ 53 ]
Applying Machine Learning to Structured Data
This could happen for a number of reasons including that there might be another
fraud prevention system in place that blocks the transactions, or that the origin
account for the transaction has insufficient funds.
out: 0.8966412742382271
As we can see in the output, close to 90% of the odd transactions have insufficient funds
in their origin accounts. From this, we can now construct a rationale in which fraudsters
try to drain a bank account of all its funds more often than regular people do.
We need this rationale to avoid the data trap. Once established, the rationale
must be constantly scrutinized. In our case, it has failed to explain 10% of the odd
transactions, and if this number rises, it could end up hurting the performance of our
model in production.
Therefore, in this section, we're going to work on preparing the data for Keras so that
we can meaningfully work with it.
Before we start, let's look at the three types of data, Nominal, Ordinal, and Numerical:
[ 54 ]
Chapter 2
Both nominal and ordinal data are categorical data, as they describe discrete
categories. While numerical data works fine with neural networks, only out of the
box, categorical data needs special treatment.
One-hot encoding
The most commonly used method to encode categorical data is called one-hot
encoding. In one-hot encoding, we create a new variable, a so-called dummy
variable for each category. We then set the dummy variable to 1 if the transaction
is a member of a certain category and to zero otherwise.
An example of how we could apply this to our data set can be seen as follows:
So, this is what the categorical data would look like before one-hot encoding:
Transaction Type
1 TRANSFER
2 CASH_OUT
3 TRANSFER
This is what the data would look like after one-hot encoding:
The Pandas software library offers a function that allows us to create dummy
variables out of the box. Before doing so, however, it makes sense to add Type_
in front of all actual transaction types. The dummy variables will be named after
the category. By adding Type_ to the beginning, we know that these dummy
variables indicate the type.
[ 55 ]
Applying Machine Learning to Structured Data
We can now get the dummy variables by running the following code:
dummies = pd.get_dummies(df['type'])
We should note that the get_dummies() function creates a new data frame. Next
we attach this data frame to the main data frame, which can be done by running:
df = pd.concat([df,dummies],axis=1)
The concat() method, as seen in the preceding code, concatenates two data frames.
We concatenate along axis 1 to add the data frame as new columns. Now that the
dummy variables are in our main data frame, we can remove the original column
by running this:
del df['type']
And, voilà! We have turned our categorical variable into something a neural network
will be able to work with.
Entity embeddings
In this section, we're going to walk through making use of both embeddings and the
Keras functional API, showing you the general workflow. Both of these topics get
introduced and explored fully in Chapter 5, Parsing Textual Data with Natural Language
Processing, where we will go beyond the general ideas presented here and where
we'll begin discussing topics like implementation.
It's fine if you do not understand everything that is going on just now; this is an
advanced section after all. If you want to use both of these techniques, you will
be well prepared after reading this book, as we explain different elements of both
methods throughout the book.
In this section, we will be creating embedding vectors for categorical data. Before
we start, we need to understand that embedding vectors are vectors representing
categorical values. We use embedding vectors as inputs for neural networks.
We train embeddings together with a neural network, so that we can, over time,
obtain more useful embeddings. Embeddings are an extremely useful tool to have
at our disposal.
Why are embeddings so useful? Not only do embeddings reduce the number of
dimensions needed for encoding over one-hot encoding and thus decrease memory
usage, but they also reduce sparsity in input activations, which helps reduce
overfitting, and they can encode semantic meanings as vectors. The same advantages
that made embeddings useful for text, Chapter 5, Parsing Textual Data with Natural
Language Processing, also make them useful for categorical data.
[ 56 ]
Chapter 2
Tokenizing categories
Just as with text, we have to tokenize the inputs before feeding them into the
embeddings layer. To do this, we have to create a mapping dictionary that maps
categories to a token. We can achieve this by running:
map_dict = {}
for token, value in enumerate(df['type'].unique()):
map_dict[value] = token
This code loops over all the unique type categories while counting upward. The
first category gets token 0, the second 1, and so on. Our map_dict looks like this:
{'CASH_IN': 4, 'CASH_OUT': 2, 'DEBIT': 3, 'PAYMENT': 0,
'TRANSFER': 1}
We have to deal with the non-categorical values in our data frame separately. We can
create a list of columns that are not the type and not the target like this:
other_cols = [c for c in df.columns if ((c != 'type') and (c !=
'isFraud'))]
The model that acts as an input for the type receives a one-dimensional input and
parses it through an embedding layer. The outputs of the embedding layer are then
reshaped into flat arrays, as we can see in this code:
num_types = len(df['type'].unique())
type_embedding_dim = 3
type_in = Input(shape=(1,))
type_embedding = Embedding(num_types,type_embedding_dim,input_
length=1)(type_in)
[ 57 ]
Applying Machine Learning to Structured Data
type_out = Reshape(target_shape=
(type_embedding_dim,))(type_embedding)
type_model = Model(type_in,type_out)
inputs.append(type_in)
outputs.append(type_out)
The type embeddings have three layers here. This is an arbitrary choice, and
experimentation with different numbers of dimensions could improve the results.
For all the other inputs, we create another input that has as many dimensions as
there are non-categorical variables and consists of a single dense layer with no
activation function. The dense layer is optional; the inputs could also be directly
passed into the head model. More layers could also be added, including these:
num_rest = len(other_cols)
rest_model = Model(rest_in,rest_out)
inputs.append(rest_in)
outputs.append(rest_out)
Now that we have created the two input models, we can concatenate them. On top
of the two concatenated inputs, we will also build our head model. To begin this
process, we must first run the following:
concatenated = Concatenate()(outputs)
Then, by running the following code, we can build and compile the overall model:
x = Dense(16)(concatenated)
x = Activation('sigmoid')(x)
x = Dense(1)(concatenated)
model_out = Activation('sigmoid')(x)
[ 58 ]
Chapter 2
Then, we can train the model by providing a list of the two inputs and the target,
as we can see in the following code:
history = merged_model.fit([types.values,rest.values],target.values,
epochs = 1,
batch_size = 128)
out:
Epoch 1/1
6362620/6362620 [==============================] - 78s 12us/step - loss:
0.0208 - acc: 0.9987
Now that we've got the columns, our data is prepared, and we can use it to create
a model.
[ 59 ]
Applying Machine Learning to Structured Data
The first step only returns the isFraud column and assigns it to y_df.
The second step returns all columns except isFraud and assigns them to x_df.
We also need to convert our data from a pandas DataFrame to NumPy arrays. The
pandas DataFrame is built on top of NumPy arrays but comes with lots of extra
bells and whistles that make all the preprocessing we did earlier possible. To train
a neural network, however, we just need the underlying data, which we can get by
simply running the following:
y = y_df.values
X = x_df.values
A test set is a holdout dataset, which we only use to evaluate our model once we
think it is doing fairly well in order to see how well it performs on data it has not
seen yet. A test set is usually randomly sampled from the complete data. Scikit-learn
offers a convenient function to do this, as we can see in the following code:
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y,
test_size=0.33, random_state=42)
The element, train_test_split will randomly assign rows to either the train or test
set. You can specify test_size, the share of data that goes into the test set (which
in our case is 33%), as well as a random state. Assigning random_state makes sure
that although the process is pseudo-random, it will always return the same split,
which makes our work more reproducible. Note that the actual choice of number
(for example, 42) does not really matter. What matters is that the same number is
used in all experiments.
[ 60 ]
Chapter 2
The answer is that every time you evaluate on the test set, you incur a bit of
"information leakage," that is, information from the test set leaks into your model
by influencing your choice of model. Gradually, the test set becomes less valuable.
The validation set is a sort of a "dirty test set" that you can use to frequently test
your models out of sample performance without worrying. Though it's key to note
that we don't want to use the test set too often, but it is still used to measure out-of-
sample performance frequently.
To this end, we'll create a "validation set," also known as a development set.
We can do this the same way we created the test set, by just splitting the training
data again, as we can see in the following code:
X_train, X_test, y_train, y_test = train_test_split(X_train,
y_train, test_size=0.1, random_state=42)
This means that we would add data that would be fraudulent to our dataset until
we have the same amount of fraudulent transactions as genuine transactions.
Note: A useful library for this kind of task is imblearn, which includes
a SMOTE function. See, http://contrib.scikitlearn.org/
imbalanced-learn/.
[ 61 ]
Applying Machine Learning to Structured Data
In practice, many structured data problems require very low learning rates. To set
the learning rate for the gradient descent optimizer, we also need to import the
optimizer. We can do this by running:
from keras.optimizers import SGD
You can see here a logistic regressor, which is the same as a one-layer neural
network:
model.compile(loss='binary_crossentropy',
optimizer=SGD(lr=1e-5),
metrics=['acc'])
Here, we will compile the model. Instead of just passing SGD to specify the optimizer
for Stochastic Gradient Descent, we'll create a custom instance of SGD in which we
set the learning rate to 0.00001. In this example, tracking accuracy is not needed
since we evaluate our models using the F1 score. Still, it still reveals some interesting
behavior, as you can see in the following code:
model.fit(X_train_res,y_train_res,
epochs=5,
batch_size=256,
validation_data=(X_val,y_val))
[ 62 ]
Chapter 2
Notice how we have passed the validation data into Keras by creating a tuple in
which we store data and labels. We will train this model for 5 epochs:
Train on 3331258 samples, validate on 185618 samples Epoch 1/5
3331258/3331258 [==============================] - 20s 6us/step - loss:
3.3568 - acc: 0.7900 - val_loss: 3.4959 - val_acc: 0.7807 Epoch 2/5
3331258/3331258 [==============================] - 20s 6us/step - loss:
3.0356 - acc: 0.8103 - val_loss: 2.9473 - val_acc: 0.8151 Epoch 3/5
3331258/3331258 [==============================] - 20s 6us/step - loss:
2.4450 - acc: 0.8475 - val_loss: 0.9431 - val_acc: 0.9408 Epoch 4/5
3331258/3331258 [==============================] - 20s 6us/step - loss:
2.3416 - acc: 0.8541 - val_loss: 1.0552 - val_acc: 0.9338 Epoch 5/5
3331258/3331258 [==============================] - 20s 6us/step - loss:
2.3336 - acc: 0.8546 - val_loss: 0.8829 - val_acc: 0.9446
Notice a few things here: first, we have trained on about 3.3 million samples,
which is more data than we initially had. The sudden increase comes from the
oversampling that we did earlier on in this chapter. Secondly, the training set's
accuracy is significantly lower than the validation set's accuracy. This is because the
training set is balanced, while the validation set is not.
We oversampled the data by adding more fraud cases to the training set than there
are in real life, which as we discussed, helped our model detect fraud better. If we
did not oversample, our model would be inclined to classify all transactions as
genuine since the vast majority of samples in the training set are genuine.
By adding fraud cases, we are forcing the model to learn what distinguishes a fraud
case. Yet, we want to validate our model on realistic data. Therefore, our validation
set does not artificially contain many fraud cases.
A model classifying everything as genuine would have over 99% accuracy on the
validation set, but just 50% accuracy on the training set. Accuracy is a flawed metric
for such imbalanced datasets. It is a half-decent proxy and more interpretable than
just a loss, which is why we keep track of it in Keras.
To evaluate our model, we should use the F1 score that we discussed at the
beginning of this chapter. However, Keras is unable to directly track the F1 score
in training since the calculation of an F1 score is somewhat slow and would end up
slowing down the training of our model.
[ 63 ]
Applying Machine Learning to Structured Data
Compare the training set's performance against the previous training set's
performance, and likewise the validation set's performance against the previous
validation set's performance. However, be careful when comparing the training set's
performance to that of the validation set's performance on highly imbalanced data.
However, if your data is equally balanced, then comparing the validation set and the
training set is a good way to gauge overfitting.
We are now in a position where we can make predictions on our test set in order to
evaluate the baseline. We start by using model.predict to make predictions on the
test set:
y_pred = model.predict(X_test)
Before evaluating our baseline, we need to turn the probabilities given by our model
into absolute predictions. In our example, we'll classify everything that has a fraud
probability above 50% as fraud. To do this, we need to run the following code:
y_pred[y_pred > 0.5] = 1
y_pred[y_pred < 0.5] = 0
Our F1 score is already significantly better than it was for the heuristic model,
which if you go back, you'll see that it only achieved a rate of 0.013131315551742895:
f1_score(y_pred=y_pred,y_true=y_test)
out: 0.054384286716408395
By plotting the confusion matrix, we're able to see that our feature-based model has
indeed improved on the heuristic model:
cm = confusion_matrix(y_pred=y_pred,y_true=y_test)
plot_confusion_matrix(cm,['Genuine','Fraud'], normalize=False)
[ 64 ]
Chapter 2
But what if we wanted to build more complex models that can express more subtle
relationships, than the one that we've just built? Let's now do that!
model.compile(loss='binary_crossentropy',
optimizer=SGD(lr=1e-5),
metrics=['acc'])
model.fit(X_train_res,y_train_res,
epochs=5, batch_size=256,
validation_data=(X_val,y_val))
y_pred = model.predict(X_test)
y_pred[y_pred > 0.5] = 1
y_pred[y_pred < 0.5] = 0
[ 65 ]
Applying Machine Learning to Structured Data
After running that code, we'll then again benchmark with the F1 score:
f1_score(y_pred=y_pred,y_true=y_test)
out: 0.087220701988752675
In this case, the more complex model does better than the simple baseline created
earlier. It seems as though the function mapping transaction data to fraud is complex
and can be approximated better by a deeper network.
In this section we have built and evaluated both simple and complex neural network
models for fraud detection. We have been careful to use the validation set to gauge
the initial out-of-sample performance.
With all of that, we can build much more complex neural networks (and we will).
But first we will have a look at the workhorse of modern enterprise-ready machine
learning: tree-based methods.
Note: For a deeper dive into XGBoost, check out the tutorials on the
XGBoost documentation page: http://xgboost.readthedocs.
io. There is a nice explanation of how tree-based methods and gradient
boosting work in theory and practice under the Tutorials section of the
website.
Let's assume for a second that our isNight feature is the greatest predictor
of fraud. A decision tree would split our dataset according to whether the
transactions happened at night or not. It would look at all the night-time
transactions, looking for the next best predictor of fraud, and it would do
the same for all day-time transactions.
[ 66 ]
Chapter 2
Scikit-learn has a handy decision tree module. We can create one for our data by
simply running the following code:
from sklearn.tree import DecisionTreeClassifier
dtree=DecisionTreeClassifier()
dtree.fit(X_train,y_train)
Simple decision trees, like the one we've produced, can give a lot of insight into data.
For example, in our decision tree, the most important feature seems to be the old
balance of the origin account, given that it is the first node in the tree.
A random forest
A more advanced version of a simple decision tree is a random forest, which is a
collection of decision trees. A forest is trained by taking subsets of the training data
and training decision trees on those subsets.
Often, those subsets do not include every feature of the training data. By doing it
this way, the different decision trees can fit different aspects of the data and capture
more information on aggregate. After a number of trees have been created, their
predictions are averaged to create the final prediction.
The idea is that the errors presented by the trees are not correlated, and so by using
multiple trees you cancel out the error. You can create and train a random forest
classifier like this:
from sklearn.ensemble import RandomForestClassifier
rf = RandomForestClassifier(n_estimators=10,n_jobs=-1)
rf.fit(X_train_res,y_train_res)
[ 67 ]
Applying Machine Learning to Structured Data
You'll notice that with the code we've just generated, random forests have far fewer
knobs to tune than neural networks. In this case, we just specify the number of
estimators, that is, the number of trees we would like our forest to have.
The n_jobs argument tells the random forest how many trees we would like to train
in parallel. Note that -1 stands for "as many as there are CPU cores":
y_pred = rf.predict(X_test)
f1_score(y_pred=y_pred,y_true=y_test)
out: 0.8749502190362406
The random forest does an order of magnitude better than the neural network as its
F1 score is close to 1, which is the maximum score. Its confusion plot, seen as follows,
shows that the random forest significantly reduced the number of false positives:
A shallow learning approach, such as a random forest, often does better than deep
learning on relatively simple problems. The reason for this is that simple relationships
with low-dimensional data can be hard to learn for a deep learning model, which has
to fit multiple parameters exactly in order to match the simple function.
As we will see in later chapters of this book, as soon as relationships do get more
complex, deep learning gets to shine.
[ 68 ]
Chapter 2
XGBoost
XGBoost stands for eXtreme Gradient Boosting. The idea behind gradient boosting
is to train a decision tree, and then to train a second decision tree on the errors that
the first decision tree made.
Through this method, multiple layers of decision trees can be added, which
slowly reduces the total number of model errors. XGBoost is a popular library
that implements gradient boosting very efficiently.
Gradient boosting classifiers can be created and trained just like random forests from
sklearn, as can be seen in the following code:
out: 0.85572959604286891
The gradient booster performs at almost the same level as a random forest on this
task. A common approach that is used is to take both a random forest and a gradient
booster and to average the predictions in order to get an even better model.
The bulk of machine learning jobs in business today are done on relatively
simple structured data. The methods we have learned today, random forests and
gradient boosting, are therefore the standard tools that most practitioners use in
the real world.
In most enterprise machine learning applications, value creation does not come
from carefully tweaking a model or coming up with cool architectures, but from
massaging data and creating good features. However, as tasks get more complex
and more semantic understanding of unstructured data is needed, these tools
begin to fail.
[ 69 ]
Applying Machine Learning to Structured Data
E2E modeling
Our current approach relies on engineered features. As we discussed at the start of
this chapter, an alternative method is E2E modeling. In E2E modeling, both raw and
unstructured data about a transaction is used. This could include the description text
of a transfer, video feeds from cameras monitoring a cash machine, or other sources
of data. E2E is often more successful than feature engineering, provided that you
have enough data available.
To get valid results, and to successfully train the data with an E2E model it can take
millions of examples. Yet, often this is the only way to gain an acceptable result,
especially when it is hard to codify the rules for something. Humans can recognize
things in images well, but it is hard to come up with exact rules that distinguish
things, which is where E2E shines.
In the dataset used for this chapter, we do not have access to more data, but the rest
of the chapters of this book demonstrate various E2E models.
Exercises
If you visit https://kaggle.com, search for a competition that has structured data.
One example is the Titanic competition. Here you can create a new kernel, do some
feature engineering, and try to build a predictive model.
How much can you improve it by investing time in feature engineering versus model
tweaking? Is there an E2E approach to the problem?
Summary
In this chapter, we have taken a structured data problem from raw data to
strong and reliable predictive models. We have learned about heuristic, feature
engineering, and E2E modeling. We have also seen the value of clear evaluation
metrics and baselines.
In the next chapter, we will look into a field where deep learning truly shines,
computer vision. Here, we will discover the computer vision pipeline, from working
with simple models to very deep networks augmented with powerful preprocessing
software. The ability to "see" empowers computers to enter completely new domains.
[ 70 ]
Utilizing Computer Vision
When Snapchat first introduced a filter featuring a breakdancing hotdog, the
stock price of the company surged. However, investors were less interested in
the hotdog's handstand; what actually fascinated them was that Snapchat had
successfully built a powerful form of computer vision technology.
The Snapchat app was now not only able to take pictures, but it was also able to
find the surfaces within those pictures that a hotdog could breakdance on. Their
app would then stick the hotdog there, something that could still be done when
the user moved their phone, allowing the hotdog to keep dancing in the same spot.
While the dancing hotdog may be one of the sillier applications of computer vision,
it successfully showed the world the potential of the technology. In a world full
of cameras, from the billions of smartphones, security cameras, and satellites in
use every day, to Internet of Things (IoT) devices, being able to interpret images
yields great benefits for both consumers and producers.
Computer vision allows us to both perceive and interpret the real world at scale.
You can think of it like this: no analyst could ever look at millions of satellite images
to mark mining sites and track their activity over time; it's just not possible. Yet for
computers, it's not just a possibility; it's something that's a reality here and now.
In fact, something that’s being used in the real world now, by several firms, is
retailers counting the number of cars in their parking lot in order to estimate what
the sales of goods will be in a given period.
Another important application of computer vision can be seen in finance, specifically
in the area of insurance. For instance, insurers might use drones to fly over roofs in
order to spot issues before they become an expensive problem. This could extend to
them using computer vision to inspect factories and equipment they insure.
[ 71 ]
Utilizing Computer Vision
Looking at another case in the finance sector, banks needing to comply with
Know-Your-Customer (KYC) rules are automating back-office processes and identity
verification. In financial trading, computer vision can be applied to candlestick charts
in order to find new patterns for technical analysis. We could dedicate a whole book
to the practical applications of computer vision.
In this chapter, we will be covering the building blocks of computer vision models.
This will include a focus on the following topics:
• Convolutional layers.
• Padding.
• Pooling.
• Regularization to prevent overfitting.
• Momentum-based optimization.
• Batch normalization.
• Advanced architectures for computer vision beyond classification.
• A note on libraries.
Before we start, let's have a look at all the different libraries we will be using in this
chapter:
It's worth taking a minute to note that all of these libraries, except for OpenCV,
can be installed via pip; for example, pip install keras.
OpenCV, however, will require a slightly more complex installation procedure. This
is beyond the scope of this book, but the information is well documented online via
OpenCV documentation, which you can view at the following URL: https://docs.
opencv.org/trunk/df/d65/tutorial_table_of_content_introduction.html.
Alternately, it's worth noting that both Kaggle and Google Colab come with OpenCV
preinstalled. To run the examples in this chapter, make sure you have OpenCV
installed and can import with import cv2.
[ 72 ]
Chapter 3
Filters on MNIST
What does a computer actually see when it sees an image? Well, the values of the
pixels are stored as numbers in the computer. So, when the computer sees a black-and-
white image of a seven, it actually sees something similar to the following:
[ 73 ]
Utilizing Computer Vision
The preceding is an example from the MNIST dataset. The handwritten number in
the image has been highlighted to make the figure seven visible for humans, but for
the computer, the image is really just a collection of numbers. This means we can
perform all kinds of mathematical operations on the image.
When detecting numbers, there are a few lower-level features that make a number.
For example, in this handwritten figure 7, there's a combination of one vertical
straight line, one horizontal line on the top, and one horizontal line through the
middle. In contrast, a 9 is made up of four rounded lines that form a circle at the
top and a straight, vertical line.
We're now able to present the central idea behind ConvNets. We can use small filters
that can detect a certain kind of low-level feature, such as a vertical line, and then
slide it over the entire image to detect all the vertical lines in the image.
The following screenshot shows a vertical line filter. To detect vertical lines in our
image, we need to slide this 3x3 matrix filter over the image.
Using the MNIST dataset on the following page, we start in the top-left corner and
slice out the top-left 3x3 grid of pixels, which in this case is all zeros.
We then perform an element-wise multiplication of all the elements in the filter with
all elements in the slice of the image. The nine products then get summed up, and
bias is added. This value then forms the output of the filter and gets passed on as
a new pixel to the next layer:
Z1 = ∑ A0 ∗ F1 + b1
[ 74 ]
Chapter 3
As a result, the output of our vertical line filter will look like this:
Take a minute to notice that the vertical lines are visible while the horizontal lines are
gone. Only a few artifacts remain. Also, notice how the filter captures the vertical line
from one side.
Since it responds to high pixel values on the left and low pixel values on the right,
only the right side of the output shows strong positive values. Meanwhile, the left
side of the line actually shows negative values. This is not a big problem in practice
as there are usually different filters for different kinds of lines and directions.
[ 75 ]
Utilizing Computer Vision
Using that example, we can now slide this filter over our image in the exact same
way we did with the vertical filter, resulting in the following output:
See how this filter removes the vertical lines and pretty much only leaves
the horizontal lines? The question now is what do we now pass onto the next
layer? Well, we stack the outputs of both filters on top of each other, creating
a three-dimensional cube:
[ 76 ]
Chapter 3
By adding multiple convolutional layers, our ConvNet is able to extract ever more
complex and semantic features.
[ 77 ]
Utilizing Computer Vision
This cube, which is referred to as a convolutional kernel, gets slid over the image just
like the two-dimensional matrix did before. The element-wise products then again get
summed up, the bias is added, and the outcome represents a pixel in the next layer.
Filters always capture the whole depth of the previous layer. The filters are moved
over the width and height of the image. Likewise, filters are not moved across the
depth, that is, the different channels, of an image. In technical terms, weights, the
numbers that make up the filters, are shared over width and height, but not over
different channels.
We can directly import the MNIST dataset from Keras by running the following code:
from keras.datasets import mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()
Our dataset contains 60,000 28x28-pixel images. MNIST characters are black
and white, so the data shape usually does not include channels:
x_train.shape
We will take a closer look at color channels later, but for now, let's expand our data
dimensions to show that we only have a one-color channel. We can achieve this by
running the following:
import numpy as np
x_train = np.expand_dims(x_train,-1)
x_test = np.expand_dims(x_test,-1)
x_train.shape
With the code being run, you can see that we now have a single color channel added.
Conv2D
Now we come to the meat and potatoes of ConvNets: using a convolutional layer
in Keras. Conv2D is the actual convolutional layer, with one Conv2D layer housing
several filters, as can be seen in the following code:
[ 78 ]
Chapter 3
model = Sequential()
img_shape = (28,28,1)
model.add(Conv2D(filters=6,
kernel_size=3,
strides=1,
padding='valid',
input_shape=img_shape))
When creating a new Conv2D layer, we must specify the number of filters we want
to use, and the size of each filter.
Kernel size
The size of the filter is also called kernel_size, as the individual filters are
sometimes called kernels. If we only specify a single number as the kernel size,
Keras will assume that our filters are squares. In this case, for example, our filter
would be 3x3 pixels.
Stride size
The strides parameter specifies the step size, also called the stride size, with which
the convolutional filter slides over the image, usually referred to as the feature map.
In the vast majority of cases, filters move pixel by pixel, so their stride size is set to 1.
However, there are researchers that make more extensive use of larger stride sizes in
order to reduce the spatial size of the feature map.
Like with kernel_size, Keras assumes that we use the same stride size horizontally
and vertically if we specify only one value, and in the vast majority of cases that
is correct. However, if we want to use a stride size of one horizontally, but two
vertically, we can pass a tuple to the parameter as follows: strides=(1,2). As
in the case of the filter size, this is rarely done.
[ 79 ]
Utilizing Computer Vision
Padding
Finally, we have to add padding to our convolutional layer. Padding adds zeros
around our image. This can be done if we want to prevent our feature map from
shrinking.
Let's consider a 5x5-pixel feature map and a 3x3 filter. The filter only fits on the
feature map nine times, so we'll end up with a 3x3 output. This both reduces the
amount of information that we can capture in the next feature map, and how much
the outer pixels of the input feature map can contribute to the task. The filter never
centers on them; it only goes over them once.
There are three options for padding: not using padding, known as "No" padding,
"Same" padding and "Valid" padding.
Option 1: No padding
[ 80 ]
Chapter 3
To ensure the output has the same size as the input, we can use same padding.
Keras will then add enough zeros around the input feature map so that we can
preserve the size. The default padding setting, however, is valid. This padding
does not preserve the feature map size, but only makes sure that the filter and
stride size actually fit on the input feature map:
Input shape
Keras requires us to specify the input shape. However, this is only required for the
first layer. For all the following layers, Keras will infer the input shape from the
previous layer's output shape.
The number of filters (here 6) and the filter size (here 3) are set as positional
arguments, while strides and padding default to 1 and valid respectively. If this
was a layer deeper in the network, we wouldn't even have to specify the input shape.
ReLU activation
Convolutional layers only perform a linear step. The numbers that make up
the image get multiplied with the filter, which is a linear operation.
[ 81 ]
Utilizing Computer Vision
The ReLU formula, which was used to produce the above chart, can be seen below:
ReLU(x) = max(x, 0)
In other words, the ReLU function returns the input if the input is positive. If it's not,
then it returns zero. This very simple function has been shown to be quite useful,
making gradient descent converge faster.
It is often argued that ReLU is faster because the derivative for all values above
zero is just one, and it does not become very small as the derivative for some extreme
values does, for example, with sigmoid or tanh.
ReLU is also less computationally expensive than both sigmoid and tanh. It does not
require any computationally expensive calculations, input values below zero are just
set to zero, and the rest is outputted. Unfortunately, though, ReLU activations are
a bit fragile and can "die."
When the gradient is very large and moves multiple weights towards a negative
direction, then the derivative of ReLU will also always be zero, so the weights never
get updated again. This might mean that a neuron never fires again. However,
this can be mitigated through a smaller learning rate.
[ 82 ]
Chapter 3
Because ReLU is fast and computationally cheap, it has become the default activation
function for many practitioners. To use the ReLU function in Keras, we can just name
it as the desired activation function in the activation layer, by running this code:
from keras.layers import Activation
model.add(Activation('relu'))
MaxPooling2D
It's common practice to use a pooling layer after a number of convolutional layers.
Pooling decreases the spatial size of the feature map, which in turn reduces the
number of parameters needed in a neural network and thus reduces overfitting.
Max pooling
Max pooling returns the maximum element out of a pool. This is in contrast to the
example average of AveragePooling2D, which returns the average of a pool. Max
pooling often delivers superior results to average pooling, so it is the standard most
practitioners use.
model.add(MaxPool2D(pool_size=2,
strides=None,
padding='valid'))
When using a max pooling layer in Keras, we have to specify the desired pool size.
The most common value is a 2x2 pool. Just as with the Conv2D layer, we can also
specify a stride size.
[ 83 ]
Utilizing Computer Vision
For pooling layers, the default stride size is None, in which case Keras sets the stride
size to be the same as the pool size. In other words, pools are next to each other and
don't overlap.
We can also specify padding, with valid being the default choice. However,
specifying same padding for pooling layers is extremely rare since the point
of a pooling layer is to reduce the spatial size of the feature map.
Our MaxPooling2D layer here takes 2x2-pixel pools next to each other with no
overlap and returns the maximum element. A more common way of specifying
the same layer is through the execution of the following:
model.add(MaxPool2D(2))
In this case, both strides and padding are set to their defaults, None and valid
respectively. There is usually no activation after a pooling layer since the pooling
layer does not perform a linear step.
Flatten
You might have noticed that our feature maps are three dimensional while our
desired output is a one-dimensional vector, containing the probability of each of the
10 classes. So, how do we get from 3D to 1D? Well, we Flatten our feature maps.
model.add(Flatten())
Dense
ConvNets usually consist of a feature extraction part, the convolutional layers,
as well as a classification part. The classification part is made up out of the simple
fully connected layers that we’ve already explored in Chapter 1, Neural Networks and
Gradient-Based Optimization, and Chapter 2, Applying Machine Learning to Structured Data.
To distinguish the plain layers from all other types of layers, we refer to them as
Dense layers. In a dense layer, each input neuron is connected to an output neuron.
We only have to specify the number of output neurons we would like, in this case, 10.
[ 84 ]
Chapter 3
After the linear step of the dense layer, we can add a softmax activation for multi-class
regression, just as we did in the first two chapters, by running the following code:
model.add(Activation('softmax'))
Training MNIST
Let's now put all of these elements together so we can train a ConvNet on the MNIST
dataset.
The model
First, we must specify the model, which we can do with the following code:
from keras.layers import Conv2D, Activation, MaxPool2D, Flatten, Dense
from keras.models import Sequential
img_shape = (28,28,1)
model = Sequential()
model.add(Conv2D(6,3,input_shape=img_shape))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Conv2D(12,3))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Flatten())
model.add(Dense(10))
model.add(Activation('softmax'))
[ 85 ]
Utilizing Computer Vision
In the following code, you can see the general structure of a typical ConvNet:
Conv2D
Pool
Conv2D
Pool
Flatten
Dense
The convolution and pooling layers are often used together in these blocks; you can
find neural networks that repeat the Conv2D, MaxPool2D combination tens of times.
[ 86 ]
Chapter 3
In this summary, you can clearly see how the pooling layers reduce the size of the
feature map. It's a little bit less obvious from the summary alone, but you can see
how the output of the first Conv2D layer is 26x26 pixels, while the input images are
28x28 pixels.
By using valid padding, Conv2D also reduces the size of the feature map, although
only by a small amount. The same happens for the second Conv2D layer, which
shrinks the feature map from 13x13 pixels to 11x11 pixels.
You can also see how the first convolutional layer only has 60 parameters, while
the Dense layer has 3,010, over 50 times as many parameters. Convolutional layers
usually achieve surprising feats with very few parameters, which is why they are
so popular. The total number of parameters in a network can often be significantly
reduced by convolutional and pooling layers.
out:
(60000, 28, 28)
So, we add a channel dimension with NumPy, with the following code:
import numpy as np
x_train = np.expand_dims(x_train,-1)
x_test = np.expand_dims(x_test,-1)
out:
(60000, 28, 28,1)
[ 87 ]
Utilizing Computer Vision
out:
(60000,)
It's the same as the categorical cross-entropy loss used in earlier chapters, the only
difference is that this uses sparse, that is, not one-hot encoded, targets.
Just as before, you still need to make sure that your network output has as many
dimensions as there are classes.
We're now at a point where we can compile the model, which we can do with
the following code:
model.compile(loss='sparse_categorical_crossentropy',
optimizer='adam',
metrics=['acc'])
As you can see, we are using an Adam optimizer. The exact workings of Adam are
explained in the next section, More bells and whistles for our neural network, but for now,
you can just think of it as a more sophisticated version of stochastic gradient descent.
When training, we can directly specify a validation set in Keras by running the
following code:
history = model.fit(x_train,
y_train,
batch_size=32,
epochs=5,
validation_data=(x_test,y_test))
Once we have successfully run that code, we'll get the following output:
Train on 60000 samples, validate on 10000 samples
Epoch 1/10
60000/60000 [==============================] - 19s 309us/step - loss:
5.3931 - acc: 0.6464 - val_loss: 1.9519 - val_acc: 0.8542
[ 88 ]
Chapter 3
Epoch 2/10
60000/60000 [==============================] - 18s 297us/step - loss:
0.8855 - acc: 0.9136 - val_loss: 0.1279 - val_acc: 0.9635
....
Epoch 10/10
To better see what is going on, we can plot the progress of training with the
following code:
import matplotlib.pyplot as plt
fig, ax = plt.subplots(figsize=(10,6))
gen = ax.plot(history.history['val_acc'], label='Validation
Accuracy')
fr = ax.plot(history.history['acc'],dashes=[5, 2], label='Training
Accuracy')
plt.show()
[ 89 ]
Utilizing Computer Vision
As you can see in the preceding chart, the model achieves about 98% validation
accuracy, which is pretty nice!
Momentum
In previous chapters we've explained gradient descent in terms of someone trying
to find the way down a mountain by just following the slope of the floor. Momentum
can be explained with an analogy to physics, where a ball is rolling down the same
hill. A small bump in the hill would not make the ball roll in a completely different
direction. The ball already has some momentum, meaning that its movement gets
influenced by its previous movement.
Instead of directly updating the model parameters with their gradient, we update
them with the exponentially weighted moving average. We update our parameter
with an outlier gradient, then we take the moving average, which will smoothen
out outliers and capture the general direction of the gradient, as we can see in the
following diagram:
[ 90 ]
Chapter 3
Vt = β ∗ Vt−1 + (1− β ) ∗ θt
A beta value of 0.9 would mean that 90% of the mean would come from the previous
moving average, Vt−1 , and 10% would come from the new value, θt .
Using momentum makes learning more robust against gradient descent pitfalls such
as outlier gradients, local minima, and saddle points.
We can augment the standard stochastic gradient descent optimizer in Keras with
momentum by setting a value for beta, which we do in the following code:
from keras.optimizers import SGD
momentum_optimizer = SGD(lr=0.01, momentum=0.9)
This little code snippet creates a stochastic gradient descent optimizer with a learning
rate of 0.01 and a beta value of 0.9. We can use it when we compile our model, as
we'll now do with this:
model.compile(optimizer=momentum_optimizer,
loss='sparse_categorical_crossentropy',
metrics=['acc'])
It then also computes the exponentially weighted average of the squared gradients:
S dW = β2 ∗ S dW + (1− β2 ) ∗ dW 2
[ 91 ]
Utilizing Computer Vision
This division by the root of squared gradients reduces the update speed when
gradients are very large. It also stabilizes learning as the learning algorithm does
not get thrown off track by outliers as much.
Using Adam, we have a new hyperparameter. Instead of having just one momentum
factor, β , we now have two, β1 and β2 . The recommended values for β1 and β2
are 0.9 and 0.999 respectively.
adam_optimizer=adam(lr=0.1,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08)
model.compile(optimizer=adam_optimizer,
loss='sparse_categorical_crossentropy',
metrics=['acc'])
As you have seen earlier in this chapter, we can also compile the model just by
passing the adam string as an optimizer. In this case, Keras will create an Adam
optimizer for us and choose the recommended values.
Regularization
Regularization is a technique used to avoid overfitting. Overfitting is when the
model fits the training data too well, and as a result, it does not generalize well
to either development or test data. You may see that overfitting is sometimes also
referred to as "high variance," while underfitting, obtaining poor results on training,
development, and test data, is referred to as "high bias."
In classical statistical learning, there is a lot of focus on the bias-variance tradeoff. The
argument that is made is that a model that fits very well to the training set is likely to
be overfitting and that some amount of underfitting (bias) has to be accepted in order
to obtain good outcomes. In classical statistical learning, the hyperparameters that
prevent overfitting also often prevent the training set fitting well.
[ 92 ]
Chapter 3
L2 regularization
One popular technique to counter overfitting is L2 regularization. L2 regularization
adds the sum of squared weights to the loss function. We can see an example of this
in the formula below:
λ
LRegularized (W ) = L (W ) +
2N
∑W 2
Adding this regularization to the loss function means that high weights increase
losses and the algorithm is incentivized to reduce the weights. Small weights,
those around zero, mean that the neural network will rely less on them.
Therefore, a regularized algorithm will rely less on every single feature and
every single node activation, and instead will have a more holistic view, taking
into account many features and activations. This will prevent the algorithm
from overfitting.
L1 regularization
L1 regularization is very similar to L2 regularization, but instead of adding the
sum of squares, it adds the sum of absolute values, as we can see in this formula:
λ
LRegularized (W ) = L (W )
2N
∑W
In practice, it is often a bit uncertain as to which of the two will work best, but the
difference between the two is not very large.
[ 93 ]
Utilizing Computer Vision
Regularization in Keras
In Keras, regularizers that are applied to the weights are called kernel_regularizer,
and regularizers that are applied to the bias are called bias_regularizer. You can
also apply regularization directly to the activation of the nodes to prevent them
from being activated very strongly with activity_regularizer.
For now, let's add some L2 regularization to our network. To do this, we need to run
the following code:
from keras.regularizers import l2
model = Sequential()
model.add(Conv2D(6,3,input_shape=img_shape, kernel_
regularizer=l2(0.01)))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Conv2D(12,3,activity_regularizer=l2(0.01)))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Flatten())
model.add(Dense(10,bias_regularizer=l2(0.01)))
model.add(Activation('softmax'))
In this following example, the regularizers are set to be shown off, but here they
actually harm the performance to our network. As you can see from the preceding
training results, our network is not actually overfitting, so setting regularizers harms
performance here, and as a result, the model underfits.
As we can see in the following output, in this case, the model reaches about 87%
validation accuracy:
[ 94 ]
Chapter 3
model.compile(loss='sparse_categorical_crossentropy',
optimizer = 'adam',
metrics=['acc'])
history = model.fit(x_train,
y_train,
batch_size=32,
epochs=10,
validation_data=(x_test,y_test))
You'll notice that the model achieves a higher accuracy on the validation than on the
training set; this is a clear sign of underfitting.
Dropout
As the title of the 2014 paper by Srivastava et al gives away, Dropout is A Simple Way
to Prevent Neural Networks from Overfitting. It achieves this by randomly removing
nodes from the neural network:
[ 95 ]
Utilizing Computer Vision
With dropout, each node has a small probability of having its activation set to zero.
This means that the learning algorithm can no longer rely heavily on single nodes,
much like in L2 and L1 regularization. Dropout therefore also has a regularizing effect.
In Keras, dropout is a new type of layer. It's put after the activations you want
to apply dropout to. It passes on activations, but sometimes it sets them to zero,
achieving the same effect as a dropout in the cells directly. We can see this in the
following code:
from keras.layers import Dropout
model = Sequential()
model.add(Conv2D(6,3,input_shape=img_shape))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Dropout(0.2))
model.add(Conv2D(12,3))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(10,bias_regularizer=l2(0.01)))
model.add(Activation('softmax'))
history = model.fit(x_train,
y_train,
batch_size=32,
epochs=10,
[ 96 ]
Chapter 3
validation_data=(x_test,y_test))
The low dropout value creates nice results for us, but again, the network does better
on the validation set rather than the training set, a clear sign of underfitting taking
place. Note that dropout is only applied at training time. When the model is used
for predictions, dropout doesn't do anything.
Batchnorm
Batchnorm, short for batch normalization, is a technique for "normalizing" input
data to a layer batch-wise. Each batchnorm computes the mean and standard
deviation of the data and applies a transformation so that the mean is zero and
the standard deviation is one.
This makes training easier because the loss surface becomes more "round."
Different means and standard deviations along different input dimensions
would mean that the network would have to learn a more complicated function.
In Keras, batchnorm is a new layer as well, as you can see in the following code:
from keras.layers import BatchNormalization
model = Sequential()
model.add(Conv2D(6,3,input_shape=img_shape))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(BatchNormalization())
[ 97 ]
Utilizing Computer Vision
model.add(Conv2D(12,3))
model.add(Activation('relu'))
model.add(MaxPool2D(2))
model.add(BatchNormalization())
model.add(Flatten())
model.add(Dense(10,bias_regularizer=l2(0.01)))
model.add(Activation('softmax'))
model.compile(loss='sparse_categorical_crossentropy',
optimizer = 'adam',
metrics=['acc'])
history = model.fit(x_train,
y_train,
batch_size=32,
epochs=10,
validation_data=(x_test,y_test))
Batchnorm often accelerates training by making it easier. You can see how the
accuracy rate jumps up in the first epoch here:
[ 98 ]
Chapter 3
Batchnorm also has a mildly regularizing effect. Extreme values are often overfitted
to, and batchnorm reduces extreme values, similar to activity regularization. All this
makes batchnorm an extremely popular tool in computer vision.
Therefore, we need to load the images from disk "just in time" rather than loading
them all in advance. In this section, we will be setting up an image data generator
that loads images on the fly.
We'll be using a dataset of plant seedlings in this case. This was provided by
Thomas Giselsson and others, 2017, via their publication, A Public Image Database
for Benchmark of Plant Seedling Classification Algorithms.
[ 99 ]
Utilizing Computer Vision
You may be wondering why we're looking at plants; after all, plant classifications are
not a common problem that is faced in the finance sector. The simple answer is that
this dataset lends itself to demonstrating many common computer vision techniques
and is available under an open domain license; it's therefore a great training dataset
for us to use.
Readers who wish to test their knowledge on a more relevant dataset should take
a look at the State Farm Distracted Driver dataset as well as the Planet: Understanding
the Amazon from Space dataset.
The code and data for this section and the section on stacking pretrained
models can be found and run here: https://www.kaggle.com/
jannesklaas/stacking-vgg.
Keras comes with an image data generator that can load files from disk out of the
box. To do this, you simply need to run:
from keras.preprocessing.image import ImageDataGenerator
To obtain a generator reading from the files, we first have to specify the generator.
In Keras, ImageDataGenerator offers a range of image augmentation tools, but
in our example, we will only be making use of the rescaling function.
Rescaling multiplies all values in an image with a constant. For most common
image formats, the color values range from 0 to 255, so we want to rescale by 1/255.
We can achieve this by running the following:
imgen = ImageDataGenerator(rescale=1/255)
This, however, is not yet the generator that loads the images for us. The
ImageDataGenerator class offers a range of generators that can be created
by calling functions on it.
We then have to specify the directory Keras should use, the batch size we would like,
in this case 32, as well as the target size the images should be resized to, in this case
150x150 pixels. To do this, we can simply run the following code:
train_generator = imgen.flow_from_directory('train',
batch_size=32,
target_size=(150,150))
validation_generator = imgen.flow_from_directory('validation',
batch_size=32,
tar get_size=(150,150))
[ 100 ]
Chapter 3
How did Keras find the images and how does it know which classes the images
belong to? The Keras generator expects the following folder structure:
• Root:
°° Class 0
°° img
°° img
°° …
°° Class 1
°° img
°° img
°° …
°° Class 2
°° …
Our dataset is already set up that way, and it's usually not hard to sort images
to match the generator's expectations.
Transfer learning aims to transfer the learning from one task to another task. As
humans, we are very good at transferring what we have learned. When you see a
dog that you have not seen before, you don't need to relearn everything about dogs
for this particular dog; instead, you just transfer new learning to what you already
knew about dogs. It's not economical to retrain a big network every time, as you'll
often find that there are parts of the model that we can reuse.
In this section, we will fine-tune VGG-16, originally trained on the ImageNet dataset.
The ImageNet competition is an annual computer vision competition, and the ImageNet
dataset consists of millions of images of real-world objects, from dogs to planes.
In the ImageNet competition, researchers compete to build the most accurate models.
In fact, ImageNet has driven much of the progress in computer vision over the recent
years, and the models built for ImageNet competitions are a popular basis to fine-
tune models from.
[ 101 ]
Utilizing Computer Vision
out:
Downloading data from https://github.com/fchollet/deep-learning-models/
releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5
58892288/58889256 [==============================] - 5s 0us/step
When downloading the data, we want to let Keras know that we don't want to
include the top part (the classification part); we also want to let Keras know the
desired input shape. If we do not specify the input shape, the model will accept
any image size, and it will not be possible to add Dense layers on top:
vgg_model.summary()
out:
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 150, 150, 3) 0
_________________________________________________________________
block1_conv1 (Conv2D) (None, 150, 150, 64) 1792
_________________________________________________________________
block1_conv2 (Conv2D) (None, 150, 150, 64) 36928
_________________________________________________________________
block1_pool (MaxPooling2D) (None, 75, 75, 64) 0
_________________________________________________________________
block2_conv1 (Conv2D) (None, 75, 75, 128) 73856
_________________________________________________________________
block2_conv2 (Conv2D) (None, 75, 75, 128) 147584
_________________________________________________________________
block2_pool (MaxPooling2D) (None, 37, 37, 128) 0
_________________________________________________________________
block3_conv1 (Conv2D) (None, 37, 37, 256) 295168
[ 102 ]
Chapter 3
_________________________________________________________________
block3_conv2 (Conv2D) (None, 37, 37, 256) 590080
_________________________________________________________________
block3_conv3 (Conv2D) (None, 37, 37, 256) 590080
_________________________________________________________________
block3_pool (MaxPooling2D) (None, 18, 18, 256) 0
_________________________________________________________________
block4_conv1 (Conv2D) (None, 18, 18, 512) 1180160
_________________________________________________________________
block4_conv2 (Conv2D) (None, 18, 18, 512) 2359808
_________________________________________________________________
block4_conv3 (Conv2D) (None, 18, 18, 512) 2359808
_________________________________________________________________
block4_pool (MaxPooling2D) (None, 9, 9, 512) 0
_________________________________________________________________
block5_conv1 (Conv2D) (None, 9, 9, 512) 2359808
_________________________________________________________________
block5_conv2 (Conv2D) (None, 9, 9, 512) 2359808
_________________________________________________________________
block5_conv3 (Conv2D) (None, 9, 9, 512) 2359808
_________________________________________________________________
block5_pool (MaxPooling2D) (None, 4, 4, 512) 0
=================================================================
Total params: 14,714,688
Trainable params: 14,714,688
Non-trainable params: 0
_________________________________________________________________
As you can see, the VGG model is very large, with over 14.7 million trainable
parameters. It also consists of both Conv2D and MaxPooling2D layers, both
of which we've already learned about when working on the MNIST dataset.
From this point, there are two different ways we can proceed:
[ 103 ]
Utilizing Computer Vision
Modifying VGG-16
In this section, we will be adding layers on top of the VGG-16 model, and then from
there, we will train the new, big model.
We do not want to retrain all those convolutional layers that have been trained
already, however. So, we must first "freeze" all the layers in VGG-16, which we can
do by running the following:
for the layer in vgg_model.layers:
layer.trainable = False
Keras downloads VGG as a functional API model. We will learn more about the
functional API in Chapter 6, Using Generative Models, but for now, we just want
to use the Sequential API, which allows us to stack layers through model.add().
We can convert a model with the functional API with the following code:
finetune = Sequential(layers = vgg_model.layers)
As a result of running the code, we have now created a new model called finetune
that works just like a normal Sequential model. We need to remember that
converting models with the Sequential API only works if the model can actually be
expressed in the Sequential API. Some more complex models cannot be converted.
As a result of everything we've just done, adding layers to our model is now simple:
finetune.add(Flatten())
finetune.add(Dense(12))
finetune.add(Activation('softmax'))
The newly added layers are by default trainable, while the reused model socket
is not. We can train this stacked model just as we would train any other model,
on the data generator we defined in the previous section. This can be executed by
running the following code:
finetune.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics = ['acc'])
finetune.fit_generator(train_generator,
epochs=8,
steps_per_epoch= 4606 // 32,
validation_data=validation_generator,
validation_steps= 144//32)
After running this, the model manages to achieve a rate of about 75% validation
accuracy.
[ 104 ]
Chapter 3
The idea is to get a large number of different images out of one image, therefore
reducing the chance that the model will overfit. For most image augmentation
purposes, we can just use Keras' ImageDataGenerator.
More advanced augmentations can be done with the OpenCV library. However,
focusing on this is outside the scope of this chapter.
This is different from rule-based augmentation, where we try to create images that
are easier to classify. For this reason, we need to create two ImageDataGenerator
instances, one for training and one for validation. This can be done by running the
following code:
train_datagen = ImageDataGenerator(
rescale = 1/255,
rotation_range=90,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.2,
zoom_range=0.1,
horizontal_flip=True,
fill_mode='nearest')
This training data generator makes use of a few built-in augmentation techniques.
Note: There are more commands available in Keras. For a full list, you
should refer to the Keras documentation at https://keras.io/.
[ 105 ]
Utilizing Computer Vision
We can check out what the generator does by running one image through it
multiple times.
First, we need to import the Keras image tools and specify an image path (this one
was chosen at random). This can be done by running the following:
from keras.preprocessing import image
fname = 'train/Charlock/270209308.png'
We then need to load the image and convert it to a NumPy array, which is achieved
with the following code:
img = image.load_img(fname, target_size=(150, 150))
img = image.img_to_array(img)
We then use the ImageDataGenerator instance we just created, but instead of using
flow_from_directory, we'll use flow, which allows us to pass the data directly
into the generator. We then pass that one image we want to use, which we can do
by running this:
gen = train_datagen.flow(img, batch_size=1)
[ 106 ]
Chapter 3
imgplot = plt.imshow(image.array_to_img(batch[0]))
plt.show()
While we could have built a model to locate and classify the plants for us, in addition
to classifying it, we could have also built a system that would output the treatment
a plant should directly receive. This begs the question of how modular we should
make our systems.
End-to-end deep learning was all the rage for several years. If given a huge amount
of data, a deep learning model can learn what would otherwise have taken a system
with many components much longer to learn. However, end-to-end deep learning
does have several drawbacks:
[ 107 ]
Utilizing Computer Vision
Recently, researchers have begun to make their models more modular. A great
example is Ha and Schmidthuber's World Models, which can be read here: https://
worldmodels.github.io/. In this, they've encoded visual information, made
predictions about the future, and chosen actions with three different models.
On the practical side, we can take a look at Airbnb, who combine structural
modeling with machine learning for their pricing engine. You can read more about it
here: https://medium.com/airbnb-engineering/learning-market-dynamics-
for-optimal-pricing-97cffbcc53e3. Modelers knew that bookings roughly
follow a Poisson Distribution and that there are also seasonal effects. So, Airbnb built
a model to predict the parameters of the distribution and seasonality directly, rather
than letting the model predict bookings directly.
If you have a small amount of data, then your algorithm's performance needs to
come from human insight. If some subtasks can be easily expressed in code, then
it's usually better to express them in code. If you need explainability and want to
see why certain choices were made, a modular setup with clearly interpretable
intermediate outputs is a good choice. However, if a task is hard and you don't know
exactly what subtasks it entails, and you have lots of data, then it's often better to use
an end-to-end approach.
It's very rare to use a pure end-to-end approach. Images, for example, are always
preprocessed from the camera chip, you never really work with raw data.
Being smart about dividing a task can boost performance and reduce risk.
In this final section of the chapter, we will discuss some approaches that go beyond
classifying images. These tasks often require more creative use of neural networks
than what we've discussed throughout this chapter.
To get the most out of this section, you don't need to worry too much about the
details of the techniques presented, but instead look at how researchers were creative
about using neural networks. We're taking this approach because you will often find
that the tasks you are looking to solve require similar creativity.
[ 108 ]
Chapter 3
Facial recognition
Facial recognition has many applications for retail institutions. For instance,
if you're in the front office, you might want to automatically recognize your
customer at an ATM, or alternatively, you might want to offer face-based security
features, such as the iPhone offers. In the back office, however, you need to comply
with KYC regulations, which require you to identify which customer you are
working with.
On the surface, facial recognition looks like a classification task. You give an image
of a face to the machine, and it will predict which person it is. The trouble is that
you might have millions of customers, but only one or two pictures per customer.
On top of that, you'll likely be continuously getting new customers. You can't
change your model every time you get a new customer, and a simple classification
approach will fail if it has to choose between millions of classes with only one
example for each class.
The creative insight here is that instead of classifying the customer's face, you can see
whether two images show the same face. You can see a visual representation of this
idea in the following diagram:
To this end, you'll have to run the two images through first. A Siamese network is a
class of neural network architecture that contains two or more identical subnetworks,
both of which are identical and contain the same weights. In Keras, you can achieve
such a setup by defining the layers first and then using them in both networks. The
two networks then feed into a single classification layer, which determines whether
the two images show the same face.
[ 109 ]
Utilizing Computer Vision
To avoid running all of the customer images in our database through the entire
Siamese network every time we want to recognize a face, it's common to save the
final output of the Siamese network. The final output of the Siamese network for
an image is called the face embedding. When we want to recognize a customer,
we compare the embedding of the image of the customer's face with the embeddings
stored in our database. We can do this with a single classification layer.
Storing facial embedding is very beneficial as it will save us a significant amount
of computational cost, in addition to allowing for the clustering of faces. Faces
will cluster together according to traits such as sex, age, and race. By only comparing
an image to the images in the same cluster, we can save even more computational
power and, as a result, get even faster recognition.
There are two ways to train Siamese networks. We can train them together with the
classifier by creating pairs of matching and non-matching images and then using
binary cross-entropy classification loss to train the entire model. However, another,
and in many respects better, option is to train the model to generate face embeddings
directly. This approach is described in Schroff, Kalenichenko, and Philbin's 2015
paper, FaceNet: A Unified Embedding for Face Recognition and Clustering, which you
can read here: https://arxiv.org/abs/1503.03832.
The idea is to create triplets of images: one anchor image, one positive image
showing the same face as the anchor image, and one negative image showing
a different face than the anchor image. A triplet loss is used to make the distance
between the anchor's embedding and the positive's embedding smaller, and the
distance between the anchor and the negative larger.
N
2 2
L = ∑ f ( xia ) − f ( xip ) − f ( xia ) − f ( xin ) + α
i
2 2
Here xia is an anchor image, and f ( xia ) is the output of the Siamese network, the
anchor image's embedding. The triplet loss is the Euclidean distance between the
anchor and the positive minus the Euclidean distance between the anchor and the
negative. A small constant, α , is a margin enforced between positive and negative
pairs. To reach zero loss, the difference between distances needs to be α .
[ 110 ]
Chapter 3
You should be able to understand that you can use a neural network to predict
whether two items are semantically the same in order to get around large
classification problems. You can train the Siamese model through some binary
classification tasks but also by treating the outputs as embeddings and using
a triplet loss. This insight extends to more than faces. If you wanted to compare
time series to classify events, then you could use the exact same approach.
What you need are bounding box predictions. A bounding box predictor outputs the
coordinates of several bounding boxes together with predictions for what object is
shown in the box.
[ 111 ]
Utilizing Computer Vision
The approach has been refined, culminating in Ren and others' 2016 paper, Faster
R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, which is
available at https://arxiv.org/abs/1506.01497, but the basic concept of sliding
a classifier over an image has remained the same.
You Only Look Once (YOLO), on the other hand, uses a single model consisting of
only convolutional layers. It divides an image into a grid and predicts an object class
for each grid cell. It then predicts several possible bounding boxes containing objects
for each grid cell.
For each bounding box, it regresses coordinates and both width and height values,
as well as a confidence score that this bounding box actually contains an object. It
then eliminates all bounding boxes with a too low confidence score or with a too
large overlap with another, a more confident bounding box.
For a more detailed description, read Redmon and Farhadi's 2016 paper,
YOLO9000: Better, Faster, Stronger, available at https://arxiv.org/
abs/1612.08242. Further reading includes the 2018 paper, YOLOv3:
An Incremental Improvement. This is available at https://arxiv.org/
abs/1804.027.
Both are well-written, tongue-in-cheek papers, that explain the YOLO
concept in more detail.
The main advantage of YOLO over an R-CNN is that it's much faster. Not having to
slide a large classification model is much more efficient. However, an R-CNN's main
advantage is that it is somewhat more accurate than a YOLO model. If your task
requires real-time analysis, you should use YOLO; however, if you do not need real-
time speed but just want the best accuracy, then using an R-CNN is the way to go.
Bounding box detection is often used as one of many processing steps. In the
insurance case, the bounding box detector would crop out all roofs. The roof images
can then be judged by a human expert, or by a separate deep learning model that
classifies damaged roofs. Of course, you could train an object locator to distinguish
between damaged and intact roofs directly, but in practice, this is usually not
a good idea.
If you're interested in reading more about this, Chapter 4, Understanding Time Series,
has a great discussion on modularity.
[ 112 ]
Chapter 3
Exercises
Fashion MNIST is a drop-in replacement for MNIST, but instead of handwritten
digits, it is about classifying clothes. Try out the techniques we have used in this
chapter on Fashion MNIST. How do they work together? What gives good results?
You can find the dataset on Kaggle at https://www.kaggle.com/zalando-
research/fashionmnist.
Take on the whale recognition challenge and read the top kernels and discussion posts.
The link can be found here: https://www.kaggle.com/c/whale-categorization-
playground. The task of recognizing whales by their fluke is similar to recognizing
humans by their face. There are good kernels showing off bounding boxes as well as
Siamese networks. We have not covered all the technical tools needed to solve the task
yet, so do not worry about the code in detail but instead focus on the concepts shown.
Summary
In this chapter, you have seen the building blocks of computer vision models. We've
learned about convolutional layers, and both the ReLU activation and regularization
methods. You have also seen a number of ways to use neural networks creatively,
such as with Siamese networks and bounding box predictors.
You have also successfully implemented and tested all these approaches on
a simple benchmark task, the MNIST dataset. We scaled up our training and used
a pretrained VGG model to classify thousands of plant images, before then using
a Keras generator to load images from disk on the fly and customizing the VGG
model to fit our new task.
We also learned about the importance of image augmentation and the modularity
tradeoff in building computer vision models. Many of these building blocks,
such as convolutions, batchnorm, and dropout, are used in other areas beyond
computer vision. They are fundamental tools that you will see outside of computer
vision applications as well. By learning about them here, you have set yourself up
to discover a wide range of possibilities in, for example, time series or generative
models.
[ 113 ]
Utilizing Computer Vision
Over the course of this chapter, we've seen that an entire pipeline is involved in a
successful computer vision project, and working on the pipeline often has a similar
or greater benefit as compared to working on the model.
In the next chapter, we will look at the most iconic and common form of financial
data: time series. We will tackle the task of forecasting web traffic using more
traditional statistical methods, such as ARIMA (short for AutoRegressive
Integrated Moving Average), as well as modern neural network-based approaches.
You will also learn about feature engineering with autocorrelation and Fourier
transformations. Finally, you will learn how to compare and contrast different
forecasting methods and build a high-quality forecasting system.
[ 114 ]
Understanding Time Series
A time series is a form of data that has a temporal dimension and is easily the
most iconic form of financial data out there. While a single stock quote is not
a time series, take the quotes you get every day and line them up, and you get
a much more interesting time series. Virtually all media materials related to finance
sooner or later show a stock price gap; not a list of prices at a given moment, but
a development of prices over time.
You'll often hear financial commenters discussing the movement of prices: "Apple
Inc. is up 5%." But what does that mean? You'll hear absolute values a lot less, such
as, "A share of Apple Inc. is $137.74." Again, what does that mean? This occurs
because market participants are interested in how things will develop in the future
and they try to extrapolate these forecasts from how things developed in the past:
[ 115 ]
Understanding Time Series
In this chapter, we will be looking at a few classic tools that are still very much
relevant today. We will then learn how neural networks can deal with time series,
and how deep learning models can express uncertainty.
Before we jump into looking at time series, I need to set your expectations for this
chapter. Many of you might have come to this chapter to read about stock market
forecasting, but I need to warn you that this chapter is not about stock market
forecasting, neither is any other chapter in this book.
Economic theory shows that markets are somewhat efficient. The efficient
market hypothesis states that all publicly available information is included in
stock prices. This extends to information on how to process information, such
as forecasting algorithms.
If this book were to present an algorithm that could predict prices on the stock
market and deliver superior returns, many investors would simply implement
this algorithm. Since those algorithms would all buy or sell in anticipation of price
changes, they would change the prices in the present, thus destroying the advantage
that you would gain by using the algorithm. Therefore, the algorithm presented
would not work for future readers.
Instead, this chapter will use traffic data from Wikipedia. Our goal is to forecast
traffic for a specific Wikipedia page. We can obtain the Wikipedia traffic data via
the wikipediatrend CRAN package.
The dataset that we are going to use here is the traffic data of around 145,000
Wikipedia pages that has been provided by Google. The data can be obtained
from Kaggle.
[ 116 ]
Chapter 4
The data in the Page column contains the name of the page, the language of the
Wikipedia page, the type of accessing device, and the accessing agent. The other
columns contain the traffic for that page on that date.
So, in the preceding table, the first row contains the page of 2NE1, a Korean pop
band, on the Chinese version of Wikipedia, by all methods of access, but only for
agents classified as spider traffic; that is, traffic not coming from humans. While
most time series work is focused on local, time-dependent features, we can enrich
all of our models by providing access to global features.
Therefore, we want to split up the page string into smaller, more useful features.
We can achieve this by running the following code:
def parse_page(page):
x = page.split('_')
return ' '.join(x[:-3]), x[-3], x[-2], x[-1]
We split the string by underscores. The name of a page could also include an
underscore, so we separate off the last three fields and then join the rest to get
the subject of the article.
[ 117 ]
Understanding Time Series
As we can see in the following code, the third-from-last element is the sub URL, for
example, en.wikipedia.org. The second-from-last element is the access, and the
last element the agent:
parse_page(train.Page[0])
Out:
('2NE1', 'zh.wikipedia.org', 'all-access', 'spider')
When we apply this function to every page entry in the training set, we obtain a list
of tuples that we can then join together into a new DataFrame, as we can see in the
following code:
l = list(train.Page.apply(parse_page))
df = pd.DataFrame(l)
df.columns = ['Subject','Sub_Page','Access','Agent']
Finally, we must add this new DataFrame back to our original DataFrame before
removing the original page column, which we can do by running the following:
train = pd.concat([train,df],axis=1)
del train['Page']
As a result of running this code, we have successfully finished loading the dataset.
This means we can now move on to exploring it.
[ 118 ]
Chapter 4
As a result of running the previous code, we will output a bar chat that ranks the
distributions of records within our dataset:
The preceding plot shows the number of time series available for each subpage.
Wikipedia has subpages for different languages, and we can see that our dataset
contains pages from the English (en), Japanese (ja), German (de), French (fr),
Chinese (zh), Russian (ru), and Spanish (es) Wikipedia sites.
[ 119 ]
Understanding Time Series
In the bar chart we produced you may have also noted two non-country based
Wikipedia sites. Both commons.wikimedia.org and www.mediawiki.org are used
to host media files such as images.
Let's run that command again, this time focusing on the type of access:
train.Access.value_counts().plot(kind='bar')
After running this code, we'll then see the following bar chart as the output:
There are two possible access methods: mobile and desktop. There's also a third
option all-access, which combines the statistics for mobile and desktop access.
[ 120 ]
Chapter 4
There are time series available not only for spider agents, but also for all other types
of access. In classic statistical modeling, the next step would be to analyze the effect
of each of these global features and build models around them. However, this is
not necessary if there's enough data and computing power available.
If that's the case then a neural network is able to discover the effects of the global
features itself and create new features based on their interactions. There are only
two real considerations that need to be addressed for global features:
[ 121 ]
Understanding Time Series
To plot the actual page views together with a 10-day rolling mean. We can do this
by running the following code:
idx = 39457
window = 10
data = train.iloc[idx,0:-4]
name = train.iloc[idx,-4]
days = [r for r in range(data.shape[0] )]
ax.plot(days,data.values,color='grey')
ax.plot(np.convolve(data,
np.ones((window,))/window,
mode='valid'),color='black')
ax.set_yscale('log')
There is a lot going on in this code snippet, and it is worth going through it step by
step. Firstly, we define which row we want to plot. The Twenty One Pilots article is
row 39,457 in the training dataset. From there, we then define the window size for
the rolling mean.
[ 122 ]
Chapter 4
We separate the page view data and the name from the overall dataset by using the
pandas iloc tool. This allows us to index the data by row and column coordinates.
Counting the days rather than displaying all the dates of the measurements makes
the plot easier to read, therefore we are going to create a day counter for the X-axis.
Next, we set up the plot and make sure it has the desired size by setting figsize.
We also define the axis labels and the title. Next, we plot the actual page views.
Our X coordinates are the days, and the Y coordinates are the page views.
To compute the mean, we are going to use a convolve operation, which you might
be familiar with as we explored convolutions in Chapter 3, Utilizing Computer Vision.
This convolve operation creates a vector of ones divided by the window size, in this
case 10. The convolve operation slides the vector over the page view, multiplies 10-
page views with 1/10, and then sums the resulting vector up. This creates a rolling
mean with a window size 10. We plot this mean in black. Finally, we specify that we
want to use a log scale for the Y axis:
Access statistics for the Twenty One Pilots Wikipedia page with a rolling mean
[ 123 ]
Understanding Time Series
You can see there are some pretty large spikes in the Twenty One Pilots graph
we just generated, even though we used a logarithmic axis. On some days,
views skyrocket to 10 times what they were just days before. Because of that,
it quickly becomes clear that a good model will have to be able to deal with
such extreme spikes.
Before we move on, it's worth pointing out that it's also clearly visible that there
are global trends, as the page views generally increase over time.
For good measure, let's plot the interest in Twenty One Pilots for all languages.
We can do this by running the following code:
fig, ax = plt.subplots(figsize=(10, 7))
plt.ylabel('Views per Page')
plt.xlabel('Day')
plt.title('Twenty One Pilots Popularity')
ax.set_yscale('log')
idx=idx[0][0]
data = train.iloc[idx,0:-4]
handle = ax.plot(days,data.values,label=country)
ax.legend()
In this snippet, we first set up the graph, as before. We then loop over the language
codes and find the index of Twenty One Pilots. The index is an array wrapped in a
tuple, so we have to extract the integer specifying the actual index. We then extract
the page view data from the training dataset and plot the page views.
In the following chart, we can view the output of the code that we've just produced:
[ 124 ]
Chapter 4
There is clearly some correlation between the time series. The English language
version of Wikipedia (the top line) is, not surprisingly, by far the most popular.
We can also see that the time series in our datasets are clearly not stationary; they
change means and standard deviations over time.
A stationary process is one whose unconditional joint probability distribution
stays constant over time. In other words, things such as the series mean or standard
deviation should stay constant.
However, as you can see, between days 200-250 in the preceding graph, the
mean views on the page changes dramatically. This result undermines some
of the assumptions many classic modeling approaches make. Yet, financial time
series are hardly ever stationary, so it is worthwhile dealing with these problems.
By addressing these problems, we become familiar with several useful tools that
can help us handle nonstationarity.
[ 125 ]
Understanding Time Series
Mean stationarity refers to the level of a series being constant. Here, individual data
points can deviate, of course, but the long-run mean should be stable. Variance
stationarity refers to the variance from the mean being constant. Again, there may be
outliers and short sequences whose variance seems higher, but the overall variance
should be at the same level. A third kind of stationarity, which is difficult to visualize
and is not shown here, is covariance stationarity. This refers to the covariance
between different lags being constant. When people refer to covariance stationarity,
they usually mean the special condition in which mean, variance, and covariances
are stationary. Many econometric models, especially in risk management, operate
under this covariance stationarity assumption.
[ 126 ]
Chapter 4
time = np.linspace(0,10,1000)
series = time
series = series + np.random.randn(1000) *0.2
[ 127 ]
Understanding Time Series
[ 128 ]
Chapter 4
You can imagine this like the tuner on an old FM radio. As you turn the tuner,
you search through different frequencies. Every once in a while, you find a
frequency that gives you a clear signal of a particular radio station. A Fourier
transformation basically scans through the entire frequency spectrum and records
at what frequencies there is a strong signal. In terms of a time series, this is useful
when trying to find periodic patterns in the data.
Imagine that we found out that a frequency of one per week gave us a strong pattern.
This would mean that knowledge about what the traffic was ton the same day one
week ago would help our model.
When both the function and the Fourier transform are discrete, which is the
case in a series of daily measurements, it is called the discrete Fourier transform
(DFT). A very fast algorithm that is used for computing the DFT is known as the
Fast Fourier Transform (FFT), which today has become an important algorithm
in scientific computing. This theory was known to the mathematician Carl Gauss
in 1805 but was brought to light more recently by American mathematicians James
W. Cooley and John Tukey in 1965.
It's beyond the scope of this chapter to go into how and why the Fourier
transformations work, so in this section we will only be giving a brief introduction.
Imagine our function as a piece of wire. We take this wire and wrap it around a
point, and if you wrap the wire so that the number of revolutions around the point
matches the frequency of a signal, all of the signal peaks will be on one side of the
pole. This means that the center of mass of the wire will move away from the point
we wrapped the wire around.
In math, wrapping a function around a point can be achieved by multiplying the
function g(n) with e−2 πifn , where f is the frequency of wrapping, n is the number of
the item from the series, and i is the imaginary square root of -1. Readers that are not
familiar with imaginary numbers can think of them as coordinates in which each
number has a two-dimensional coordinate consisting of both a real and an imaginary
number.
To compute the center of mass, we average the coordinates of the points in our
discrete function. The DFT formula is, therefore, as follows:
N −1 fn
−2 πi
y[ f ] = ∑ e N
x [n]
n=0
[ 129 ]
Understanding Time Series
Here y[f] is the fth element in the transformed series, and x[n] is the nth element of
the input series, x. N is the total number of points in the input series. Note that y[f]
will be a number with a real and a discrete element.
To detect frequencies, we are only really interested in the overall magnitude of y[f]. To
get this magnitude we need to so we compute the root of the sum of the squares of the
imaginary and real parts. In Python, we do not have to worry about all the math as we
can use scikit-learn's fftpack, which has an FFT function built in.
The next step is to run the following code:
data = train.iloc[:,0:-4]
fft_complex = fft(data)
fft_mag = [np.sqrt(np.real(x)*np.real(x)+
np.imag(x)*np.imag(x)) for x in fft_complex]
Here, we first extract the time series measurements without the global features
from our training set. Then we run the FFT algorithm, before finally computing
the magnitudes of the transformation.
After running that code, we now have the Fourier transformations for all the time
series datasets. In order to allow us to get a better insight into the general behavior
of the Fourier transformations we can average them by simply running:
arr = np.array(fft_mag)
fft_mean = np.mean(arr,axis=0)
This first turns the magnitudes into a NumPy array before then computing the
mean. We want to compute the mean per frequency, not just the mean value of all the
magnitudes, therefore we need to specify the axis along which to take the mean value.
In this case, the series are stacked in rows, so taking the mean column-wise
(axis zero) will result in frequency-wise means. To better plot the transformation,
we need to create a list of frequencies tested. The frequencies are in the form: day/
all days in the dataset for each day, so 1/550, 2/550, 3/550, and so on. To create
the list we need to run:
fft_xvals = [day / fft_mean.shape[0] for day in
range(fft_mean.shape[0])]
[ 130 ]
Chapter 4
Upon plotting the transformation, we will have successfully produced a chart similar
to the one you see here:
As you can see in the chart we produced, there are spikes at roughly 1/7 (0.14), 2/7
(0.28), and 3/7 (0.42). As a week has seven days, that is a frequency of one time per
week, two times per week, and three times per week. In other words, page statistics
repeat themselves (approximately) every week, so that, for example, access on one
Saturday correlates with access on the previous Saturday.
Autocorrelation
Autocorrelation is the correlation between two elements of a series separated by
a given interval. Intuitively, we would, for example, assume that knowledge about
the last time step helps us in forecasting the next step. But how about knowledge
from 2 time steps ago or from 100 time steps ago?
[ 131 ]
Understanding Time Series
autocorrelation_plot(data.iloc[110])
plt.title(' '.join(train.loc[110,['Subject', 'Sub_Page']]))
The plot in the preceding chart shows the correlation of page views for the Wikipedia
page of Oh My Girl, a South Korean girl group, within the Chinese Wikipedia.
You can see that shorter time intervals between 1 and 20 days show a higher
autocorrelation than longer intervals. Likewise there are also curious spikes, such
as around 120 days and 280 days. It's possible that annual, quarterly, or monthly
events could lead to an increase in the frequency of visits to the Oh My Girl
Wikipedia page.
[ 132 ]
Chapter 4
We can examine the general pattern of these frequencies by drawing 1,000 of these
autocorrelation plots. To do this we run the following code:
a = np.random.choice(data.shape[0],1000)
for i in a:
autocorrelation_plot(data.iloc[i])
plt.title('1K Autocorrelations')
This code snippet first samples 1,000 random numbers between 0 and the number of
series in our dataset, which in our case is around 145,000. We use these as indices to
randomly sample rows from our dataset for which we then draw the autocorrelation
plot, which we can see in the following graphic:
As you can see, autocorrelations can be quite different for different series and there
is a lot of noise within the chart. There also seems to be a general trend toward
higher correlations at around the 350-day mark.
[ 133 ]
Understanding Time Series
Time series analysis, such as in the examples shown previously, can help us engineer
features for our model. Complex neural networks could, in theory, discover all of
these features by themselves. However, it is often much easier to help them a bit,
especially with information about long periods of time.
If we have few series, but multiple data observations per series, a walk-forward split
is preferable. However, if we have a lot of series, but few observations per series,
then a side-by-side split is preferable.
[ 134 ]
Chapter 4
Establishing a training and testing regime also aligns more nicely with the
forecasting problem at hand. In side-by-side splitting, the model might overfit
to global events in the prediction period. Imagine that Wikipedia was down for
a week in the prediction period used in side-by-side splitting. This event would
reduce the number of views for all the pages, and as a result the model would
overfit to this global event.
We would not catch the overfitting in our validation set because the prediction
period would also be affected by the global event. However, in our case, we
have multiple time series, but only about 550 observations per series. Therefore
there seems to be no global events that would have significantly impacted all the
Wikipedia pages in that time period.
However, there are some global events that impacted views for some pages, such
as the Winter Olympics. Yet, this is a reasonable risk in this case, as the number
of pages affected by such global events is still small. Since we have an abundance
of series and only a few observations per series, a side-by-side split is more feasible
in our case.
In this chapter, we're focusing on forecasting traffic for 50 days. So, we must first
split the last 50 days of each series from the rest, as seen in the following code,
before splitting the training and validation set:
from sklearn.model_selection import train_test_split
X = data.iloc[:,:500]
y = data.iloc[:,500:]
When splitting, we use X.values to only get the data, not a DataFrame containing
the data. After splitting we are left with 130,556 series for training and 14,507 for
validation.
In this example, we are going to use the mean absolute percentage error (MAPE)
as a loss and evaluation metric. MAPE can cause division-by-zero errors if the true
value of y is zero. Thus, to prevent division by zero occurring, we'll use a small-
value epsilon:
def mape(y_true,y_pred):
eps = 1
err = np.mean(np.abs((y_true - y_pred) / (y_true + eps))) *
100
return err
[ 135 ]
Understanding Time Series
A note on backtesting
The peculiarities of choosing training and testing sets are especially important in
both systematic investing and algorithmic trading. The main way to test trading
algorithms is a process called backtesting.
Backtesting means we train the algorithm on data from a certain time period and
then test its performance on older data. For example, we could train on data from
a date range of 2015 to 2018 and then test on data from 1990 to 2015. By doing this,
not only is the model's accuracy tested, but the backtested algorithm executes virtual
trades so its profitability can be evaluated. Backtesting is done because there is
plenty of past data available.
With all that being said, backtesting does suffer from several biases. Let's take a look
at four of the most important biases that we need to be aware of:
[ 136 ]
Chapter 4
Another option is to deploy models silently and test them in the future. The
algorithm runs but executes only virtual trades so that if things go wrong, no
money will be lost. This approach makes use of future data instead of past data.
However, the downside to this method is that we have to wait for quite a while
before the algorithm can be used.
Median forecasting
A good sanity check and an often underrated forecasting tool is medians. A median
is a value separating the higher half of a distribution from the lower half; it sits
exactly in the middle of the distribution. Medians have the advantage of removing
noise, coupled with the fact that they are less susceptible to outliers than means, and
the way they capture the midpoint of distribution means that they are also easy to
compute.
To make a forecast, we compute the median over a look-back window in our training
data. In this case, we use a window size of 50, but you could experiment with other
values. The next step is to select the last 50 values from our X values and compute
the median.
Take a minute to note that in the NumPy median function, we have to set
keepdims=True. This ensures that we keep a two-dimensional matrix rather than
a flat array, which is important when computing the error. So, to make a forecast,
we need to run the following code:
lookback = 50
lb_data = X_train[:,-lookback:]
med = np.median(lb_data,axis=1,keepdims=True)
err = mape(y_train,med)
[ 137 ]
Understanding Time Series
The output returned shows we obtain an error of about 68.1%; not bad given the
simplicity of our method. To see how the medians work, let's plot the X values, the
true y values, and predictions for a random page:
idx = 15000
fig, ax = plt.subplots(figsize=(10, 7))
ax.plot(np.arange(500),X_train[idx], label='X')
ax.plot(np.arange(500,550),y_train[idx],label='True')
ax.plot(np.arange(500,550),np.repeat(med[idx],50),
label='Forecast')
As you can see, our plotting consists of drawing three plots. For each plot, we
must specify the X and Y values for the plot. For X_train, the X values range
from 0 to 500, and for y_train and the forecast they range from 500 to 550. We
then select the series we want to plot from our training data. Since we have only
one median value, we repeat the median forecast of the desired series 50 times in
order to draw our forecast.
The output can be seen here:
Median forecast and actual values for access of an image file. The True values are to the right-hand
side of the plot, and the median forecast is the horizontal line in the center of them.
[ 138 ]
Chapter 4
As you can see in the preceding output median forecast, the data for this page, in
this case, an image of American actor Eric Stoltz, is very noisy, and the median cuts
through all the noise. The median is especially useful here for pages that are visited
infrequently and where there is no clear trend or pattern.
This is not all you can do with medians. Beyond what we've just covered, you could,
for example, use different medians for weekends or use a median of medians from
multiple look-back periods. A simple tool, such as median forecasting, is able to
deliver good results with smart feature engineering. Therefore, it makes sense
to spend a bit of time on implementing median forecasting as a baseline and
performing a sanity check before using more advanced methods.
ARIMA
Earlier, in the section on exploratory data analysis, we talked about how seasonality
and stationarity are important elements when it comes to forecasting time series.
In fact, median forecasting has trouble with both. If the mean of a time series
continuously shifts, then median forecasting will not continue the trend, and if a time
series shows cyclical behavior, then the median will not continue with the cycle.
ARIMA which stands for Autoregressive Integrated Moving Average, is made
up of three core components:
• Autoregression: The model uses the relationship between a value and
a number of lagged observations.
• Integrated: The model uses the difference between raw observations to make
the time series stationary. A time series going continuously upward will have
a flat integral as the differences between points are always the same.
• Moving Average: The model uses residual errors from a moving average.
We have to manually specify how many lagged observations we want to include, p,
how often we want to differentiate the series, d, and how large the moving average
window should be, q. ARIMA then performs a linear regression against all the included
lagged observations and moving average residuals on the differentiated series.
We can use ARIMA in Python with statsmodels, a library with many helpful
statistical tools. To do this, we simply run this:
from statsmodels.tsa.arima_model import ARIMA
Then, to create a new ARIMA model, we pass the data we want to fit, in this case
from our earlier example of views for 2NE1 from the Chinese Wikipedia, as well as
the desired values for p, d, and q, in that order. In this case, we want to include five
lagged observations, differentiate once, and take a moving average window of five.
In code, this works out as follows:
model = ARIMA(X_train[0], order=(5,1,5))
[ 139 ]
Understanding Time Series
Running model.summary() at this point would output all the coefficients as well
as significance values for statistical analysis. We, however, are more interested in
how well our model does in forecasting. So, to complete this, and see the output,
we simply run:
residuals = pd.DataFrame(model.resid)
ax.plot(residuals)
After running the previous code, we'll be able to output the results for 2NE1 page
views, as we can see in this graph:
In the preceding chart, we can see that the model does very well in the beginning
but really begins to struggle at around the 300-day mark. This could be because
page views are harder to predict or because there is more volatility in this period.
In order for us to ensure that our model is not skewed, we need to examine the
distribution of the residuals. We can do this by plotting a kernel density estimator,
which is a mathematical method designed to estimate distributions without needing
to model them.
[ 140 ]
Chapter 4
As you can see, our model roughly represents a Gaussian distribution with a mean
of zero. So, it's all good on that front, but then the question arises, "how do we make
forecasts?"
To use this model for forecasting, all we have to do is to specify the number of days
we want to forecast, which we can do with the following code:
predictions, stderr, conf_int = model.forecast(50)
This forecast not only gives us predictions but also the standard error and confidence
interval, which is 95% by default.
Let's plot the projected views against the real views to see how we are doing. This
graph shows the last 20 days for our prediction basis as well as the forecast to keep
things readable. To produce this, we must execute the following code:
[ 141 ]
Understanding Time Series
ax.plot(np.arange(480,500),basis[480:], label='X')
ax.plot(np.arange(500,550),y_train[0], label='True')
ax.plot(np.arange(500,550),predictions, label='Forecast')
You can see that ARIMA captures the periodicity of the series very well. Its forecast
does steer off a bit toward the end, but in the beginning, it does a remarkable job.
Kalman filters
Kalman filters are a method of extracting a signal from either noisy or incomplete
measurements. They were invented by Hungarian-born, American engineer, Rudolf
Emil Kalman, for the purpose of electrical engineering, and were first used in the
Apollo Space program in the 1960s.
[ 142 ]
Chapter 4
The basic idea behind the Kalman filter is that there is some hidden state of a system
that we cannot observe directly but for which we can obtain noisy measurements.
Imagine you want to measure the temperature inside a rocket engine. You cannot
put a measurement device directly into the engine, because it's too hot, but you
can have a device on the outside of the engine.
Naturally, this measurement is not going to be perfect, as there are a lot of external
factors occurring outside of the engine that make the measurement noisy. Therefore,
to estimate the temperature inside the rocket, you need a method that can deal
with the noise. We can think of the internal state in the page forecasting as the
actual interest in a certain page, of which the page views represent only a noisy
measurement.
The idea here is that the internal state, xk , at time k is a state transition matrix,
A, multiplied with the previous internal state, xk −1 , plus some process noise, qk −1 .
How interest in the Wikipedia page of 2NE1 develops is to some degree random.
The randomness is assumed to follow a Gaussian normal distribution with mean
zero and variance Q:
xk = Axk −1 + qk −1 , qk −1 ~ N (0, Q )
state_transition[0,0] = 1
[ 143 ]
Understanding Time Series
The transition matrix, A, looks like the following table, describing one hidden level,
which we might interpret as the real interest as well as a seasonality model:
array([[ 1., 0., 0., 0., 0., 0., 0., 0.],
[ 0., -1., -1., -1., -1., -1., -1., 0.],
[ 0., 1., 0., 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0., 0., 0., 0.],
[ 0., 0., 0., 1., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 1., 0., 0., 0.],
[ 0., 0., 0., 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0., 0., 1., 0.]])
The observation model, H, maps the general interest plus seasonality to a single
measurement:
observation_model = [[1,1] + [0]*(n_seasons-1)]
The noise priors are just estimates scaled by a "smoothing factor," which allows
us to control the update process:
smoothing_factor = 5.0
Otto Seiskari, a mathematician and 8th place winner in the original Wikipedia traffic
forecasting competition, wrote a very fast Kalman filtering library, which we will be
using here. His library allows for the vectorized processing of multiple independent
time series, which is very handy if you have 145,000 time series to process.
[ 144 ]
Chapter 4
From there we can then estimate the parameters and compute a forecast in one step:
result = kf.compute(X_train[0], 50)
Once again, we make forecasts for 2NE1's Chinese page and create a forecast for 50
days. Take a minute to note that we could also pass multiple series, for example, the
first 10 with X_train[:10], and compute separate filters for all of them at once.
The result of the compute function contains the state and observation estimates from
the smoothing process as well as predicted internal states and observations. States
and observations are Gaussian distributions, so to get a plottable value, we need to
access their mean.
Our states are eight-dimensional, but we only care about the non-seasonal state value,
so we need to index the mean, which we can achieve by running the following:
fig, ax = plt.subplots(figsize=(10, 7))
ax.plot(np.arange(480,500),X_train[0,480:], label='X')
ax.plot(np.arange(500,550),y_train[0],label='True')
ax.plot(np.arange(500,550),
result.predicted.observations.mean,
label='Predicted observations')
ax.plot(np.arange(500,550),
result.predicted.states.mean[:,0],
[ 145 ]
Understanding Time Series
label='predicted states')
ax.plot(np.arange(480,500),
result.smoothed.observations.mean[480:],
label='Expected Observations')
ax.plot(np.arange(480,500),
result.smoothed.states.mean[480:,0],
label='States')
ax.legend()
ax.set_yscale('log')
We can clearly see in the preceding graph the effects of our prior modeling on the
predictions. We can see the model predicts strong weekly oscillation, stronger than
actually observed. Likewise, we can also see that the model does not anticipate any
trends since we did not see model trends in our prior model.
[ 146 ]
Chapter 4
Kalman filters are a useful tool and are used in many applications, from electrical
engineering to finance. In fact, until relatively recently, they were the go-to tool
for time series modeling. Smart modelers were able to create smart systems that
described the time series very well. However, one weakness of Kalman filters is that
they cannot discover patterns by themselves and need carefully engineered priors in
order to work.
Before we can dive in and start building forecast models, we need to do some
preprocessing and feature engineering. The advantage of neural networks is that
they can take in both a high number of features in addition to very high-dimensional
data. The disadvantage is that we have to be careful about what features we input.
Remember how we discussed look-ahead bias earlier in the chapter, including future
data that would not have been available at the time of forecasting, which
is a problem in backtesting.
Data preparation
For each series, we will assemble the following features:
• log_view: The natural logarithm of page views. Since the logarithm of zero
is undefined, we will use log1p, which is the natural logarithm of page views
plus one.
• days: One-hot encoded weekdays.
• year_lag: The value of log_view from 365 days ago. -1 if there is no value
available.
• halfyear_lag: The value of log_view from 182 days ago. -1 if there is no
value available.
• quarter_lag: The value of log_view from 91 days ago. -1 if there is no
value available.
• page_enc: The one-hot encoded subpage.
• agent_enc: The one-hot encoded agent.
[ 147 ]
Understanding Time Series
These features are assembled for each time series, giving our input data the shape
(batch size, look back window size, 29).
Weekdays
The day of the week matters. Sundays may show different access behavior, when
people are browsing from their couch, compared to Mondays, when people may be
looking up things for work. So, we need to encode the weekday. A simple one-hot
encoding will do the job:
import datetime
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import OneHotEncoder
weekdays = [datetime.datetime.strptime(date, '%Y-%m
-%d').strftime('%a')
for date in train.columns.values[:-4]]
Firstly, we turn the date strings (such as 2017-03-02) into their weekday (Thursday).
This is very simple to do, and can be done with the following code:
day_one_hot = LabelEncoder().fit_transform(weekdays)
day_one_hot = day_one_hot.reshape(-1, 1)
We then encode the weekdays into integers, so that "Monday" becomes 1, "Tuesday"
becomes 2, and so on. We reshape the resulting array into a rank-2 tensor with shape
(array length, 1) so that the one-hot encoder knows that we have many observations,
but only one feature, and not the other way around:
day_one_hot =
OneHotEncoder(sparse=False).fit_transform(day_one_hot)
day_one_hot = np.expand_dims(day_one_hot,0)
Finally, we one-hot encode the days. We then add a new dimension to the tensor
showing that we only have one "row" of dates. We will later repeat the array along
this axis:
agent_int = LabelEncoder().fit(train['Agent'])
agent_enc = agent_int.transform(train['Agent'])
agent_enc = agent_enc.reshape(-1, 1)
agent_one_hot = OneHotEncoder(sparse=False).fit(agent_enc)
del agent_enc
[ 148 ]
Chapter 4
We will need the encoders for the agents later when we encode the agent of
each series.
Here, we first create a LabelEncoder instance that can transform the agent name
strings into integers. We then transform all of the agents into such an integer string
in order to set up a OneHotEncoder instance that can one-hot encode the agents. To
save memory, we will then delete the already-encoded agents.
We do the same for subpages and access methods by running the following:
page_int = LabelEncoder().fit(train['Sub_Page'])
page_enc = page_int.transform(train['Sub_Page'])
page_enc = page_enc.reshape(-1, 1)
page_one_hot = OneHotEncoder(sparse=False).fit(page_enc)
del page_enc
acc_int = LabelEncoder().fit(train['Access'])
acc_enc = acc_int.transform(train['Access'])
acc_enc = acc_enc.reshape(-1, 1)
acc_one_hot = OneHotEncoder(sparse=False).fit(acc_enc)
del acc_enc
Now we come to the lagged features. Technically, neural networks could discover
what past events are relevant for forecasting themselves. However, this is pretty
difficult because of the vanishing gradient problem, something that is covered in
more detail later, in the LSTM section of this chapter. For now, let's just set up a little
function that creates an array lagged by a number of days:
def lag_arr(arr, lag, fill):
filler = np.full((arr.shape[0],lag,1),-1)
comb = np.concatenate((filler,arr),axis=1)
result = comb[:,:arr.shape[1]]
return result
This function first creates a new array that will fill up the "empty space" from the
shift. The new array has as many rows as the original array but its series length, or
width, is the number of days we want to lag. We then attach this array to the front
of our original array. Finally, we remove elements from the back of the array in order
to get back to the original array series length or width.
We want to inform our model
about the amount of autocorrelation for different time intervals. To compute the
autocorrelation for a single series, we shift the series by the amount of lag we want
to measure the autocorrelation for. We then compute the autocorrelation:
[ 149 ]
Understanding Time Series
In this formula τ is the lag indicator. We do not just use a NumPy function since
there is a real possibility that the divider is zero. In this case, our function will just
return 0:
def single_autocorr(series, lag):
s1 = series[lag:]
s2 = series[:-lag]
ms1 = np.mean(s1)
ms2 = np.mean(s2)
ds1 = s1 - ms1
ds2 = s2 - ms2
divider = np.sqrt(np.sum(ds1 * ds1)) * np.sqrt(np.sum
(ds2 * ds2))
return np.sum(ds1 * ds2) / divider if divider != 0 else 0
We can use this function, which we wrote for a single series, to create a batch
of autocorrelation features, as seen here:
def batc_autocorr(data,lag,series_length):
corrs = []
for i in range(data.shape[0]):
c = single_autocorr(data, lag)
corrs.append(c)
corr = np.array(corrs)
corr = np.expand_dims(corr,-1)
corr = np.expand_dims(corr,-1)
corr = np.repeat(corr,series_length,axis=1)
return corr
Firstly, we calculate the autocorrelations for each series in the batch. Then we
fuse the correlations together into one NumPy array. Since autocorrelations are a
global feature, we need to create a new dimension for the length of the series and
another new dimension to show that this is only one feature. We then repeat the
autocorrelations over the entire length of the series.
The get_batch function utilizes all of these tools in order to provide us with one
batch of data, as can be seen with the following code:
def get_batch(train,start=0,lookback = 100): #1
assert((start + lookback) <= (train.shape[1] - 5)) #2
data = train.iloc[:,start:start + lookback].values #3
target = train.iloc[:,start + lookback].values
target = np.log1p(target) #4
[ 150 ]
Chapter 4
log_view = np.log1p(data)
log_view = np.expand_dims(log_view,axis=-1) #5
days = day_one_hot[:,start:start + lookback]
days = np.repeat(days,repeats=train.shape[0],axis=0) #6
year_lag = lag_arr(log_view,365,-1)
halfyear_lag = lag_arr(log_view,182,-1)
quarter_lag = lag_arr(log_view,91,-1) #7
agent_enc = agent_int.transform(train['Agent'])
agent_enc = agent_enc.reshape(-1, 1)
agent_enc = agent_one_hot.transform(agent_enc)
agent_enc = np.expand_dims(agent_enc,1)
agent_enc = np.repeat(agent_enc,lookback,axis=1) #8
page_enc = page_int.transform(train['Sub_Page'])
page_enc = page_enc.reshape(-1, 1)
page_enc = page_one_hot.transform(page_enc)
page_enc = np.expand_dims(page_enc, 1)
page_enc = np.repeat(page_enc,lookback,axis=1) #9
acc_enc = acc_int.transform(train['Access'])
acc_enc = acc_enc.reshape(-1, 1)
acc_enc = acc_one_hot.transform(acc_enc)
acc_enc = np.expand_dims(acc_enc,1)
acc_enc = np.repeat(acc_enc,lookback,axis=1) #10
year_autocorr = batc_autocorr(data,lag=365,
series_length=lookback)
halfyr_autocorr = batc_autocorr(data,lag=182,
series_length=lookback)
quarter_autocorr = batc_autocorr(data,lag=91,
series_length=lookback) #11
medians = np.median(data,axis=1)
medians = np.expand_dims(medians,-1)
medians = np.expand_dims(medians,-1)
medians = np.repeat(medians,lookback,axis=1) #12
batch = np.concatenate((log_view,
days,
year_lag,
halfyear_lag,
quarter_lag,
page_enc,
agent_enc,
acc_enc,
year_autocorr,
halfyr_autocorr,
quarter_autocorr,
medians),axis=2)
[ 151 ]
Understanding Time Series
That was a lot of code, so let's take a minute to walk through the preceding code step
by step in order to fully understand it:
1. Ensures there is enough data to create a lookback window and a target from
the given starting point.
2. Separates the lookback window from the training data.
3. Separates the target and then takes the one plus logarithm of it.
4. Takes the one plus logarithm of the lookback window and adds a feature
dimension.
5. Gets the days from the precomputed one-hot encoding of days and repeats
it for each time series in the batch.
6. Computes the lag features for year lag, half-year lag, and quarterly lag.
7. This step will encode the global features using the preceding defined
encoders. The next two steps, 8 and 9, will echo the same role.
8. This step repeats step 7.
9. This step repeats step 7 and 8.
10. Calculates the year, half-year, and quarterly autocorrelation.
11. Calculates the median for the lookback data.
12. Fuses all these features into one batch.
Finally, we can use our get_batch function to write a generator, just like we
did in Chapter 3, Utilizing Computer Vision. This generator loops over the original
training set and passes a subset into the get_batch function. It then yields the
batch obtained.
Note that we choose random starting points to make the most out of our data:
def generate_batches(train,batch_size = 32, lookback = 100):
num_samples = train.shape[0]
num_steps = train.shape[1] - 5
while True:
for i in range(num_samples // batch_size):
batch_start = i * batch_size
batch_end = batch_start + batch_size
[ 152 ]
Chapter 4
Conv1D
You might remember Convolution Neural Networks (ConvNets, or CNNs)
from Chapter 3, Utilizing Computer Vision, where we looked briefly at roofs and
insurance. In computer vision, convolutional filters slide over the image two-
dimensionally. There is also a version of convolutional filters that can slide over a
sequence one-dimensionally. The output is another sequence, much like the output
of a two-dimensional convolution was another image. Everything else about one-
dimensional convolutions is exactly the same as two-dimensional convolutions.
In this section, we're going to start by building a ConvNet that expects a fixed input
length:
n_features = 29
max_len = 100
model = Sequential()
model.add(Conv1D(16,5, input_shape=(100,29)))
model.add(Activation('relu'))
model.add(MaxPool1D(5))
model.add(Conv1D(16,5))
model.add(Activation('relu'))
model.add(MaxPool1D(5))
model.add(Flatten())
model.add(Dense(1))
Notice that next to Conv1D and Activation, there are two more layers in this
network. MaxPool1D works exactly like MaxPooling2D, which we used earlier in
the book. It takes a piece of the sequence with a specified length and returns the
maximum element in the sequence. This is similar to how it returned the maximum
element of a small window in two-dimensional convolutional networks.
Take note that max pooling always returns the maximum element for each channel.
Flatten transforms the two-dimensional sequence tensor into a one-dimensional flat
tensor. To use Flatten in combination with Dense, we need to specify the sequence
length in the input shape. Here, we set it with the max_len variable. We do this
because Dense expects a fixed input shape and Flatten will return a tensor based
on the size of its input.
An alternative to using Flatten is GlobalMaxPool1D, which returns the maximum
element of the entire sequence. Since the sequence is fixed in size, you can use
a Dense layer afterward without fixing the input length.
Our model compiles just as you would expect:
model.compile(optimizer='adam',
loss='mean_absolute_percentage_error')
[ 153 ]
Understanding Time Series
We then train it on the generator that we wrote earlier. To obtain separate train and
validation sets, we must first split the overall dataset and then create two generators
based on the two datasets. To do this, run the following code:
from sklearn.model_selection import train_test_split
batch_size = 128
train_df, val_df = train_test_split(train, test_size=0.1)
train_gen = generate_batches(train_df,batch_size=batch_size)
val_gen = generate_batches(val_df, batch_size=batch_size)
n_train_samples = train_df.shape[0]
n_val_samples = val_df.shape[0]
Finally, we can train our model on a generator, just like we did in computer vision:
model.fit_generator(train_gen,
epochs=20,
steps_per_epoch=n_train_samples // batch_size,
validation_data= val_gen,
validation_steps=n_val_samples // batch_size)
Your validation loss will still be quite high, around 12,798,928. The absolute loss
value is never a good guide for how well your model is doing. You'll find that
it's better to use other metrics in order to see whether your forecasts are useful.
However, please note that we will reduce the loss significantly later in this chapter.
Standard convolution does not take the direction of convolution into account
[ 154 ]
Chapter 4
As the convolutional filter slides over the data, it looks into the future as well as the
past. Causal convolution ensures that the output at time t derives only from inputs
from time t - 1:
In Keras, all we have to do is set the padding parameter to causal. We can do this by
executing the following code:
model.add(Conv1D(16,5, padding='causal'))
Another useful trick is dilated convolutional networks. Dilation means that the filter
only accesses every nth element, as we can see in the image below.
In the preceding diagram, the upper convolutional layer has a dilation rate of 4 and
the lower layer a dilation rate of 1. We can set the dilation rate in Keras by running
the following:
model.add(Conv1D(16,5, padding='causal', dilation_rate=4))
[ 155 ]
Understanding Time Series
Simple RNN
Another method to make order matter within neural networks is to give the network
some kind of memory. So far, all of our networks have done a forward pass without
any memory of what happened before or after the pass. It's time to change that with
a recurrent neural network (RNN):
RNNs contain recurrent layers. Recurrent layers can remember their last activation
and use it as their own input:
At = activation (W ∗ in + U ∗ At−1 + b)
A recurrent layer takes a sequence as an input. For each element, it then computes
a matrix multiplication (W * in), just like a Dense layer, and runs the result through
an activation function, such as relu. It then retains its own activation. When the
next item of the sequence arrives, it performs the matrix multiplication as before, but
this time it also multiplies its previous activation with a second matrix ( U ∗ A ). The
t −1
recurrent layer adds the result of both operations together and passes it through the
activation function again.
model = Sequential()
model.add(SimpleRNN(16,input_shape=(max_len,n_features)))
model.add(Dense(1))
model.compile(optimizer='adam',
loss='mean_absolute_percentage_error')
The only parameter we need to specify is the size of the recurrent layer. This is
basically the same as setting the size of a Dense layer, as SimpleRNN layers are very
similar to Dense layers except that they feed their output back in as input. RNNs, by
default, only return the last output of the sequence.
[ 156 ]
Chapter 4
model = Sequential()
model.add(SimpleRNN(16,return_sequences=True,
input_shape=(max_len,n_features)))
model.add(SimpleRNN(32, return_sequences = True))
model.add(SimpleRNN(64))
model.add(Dense(1))
model.compile(optimizer='adam',
loss='mean_absolute_percentage_error')
model.fit_generator(train_gen,
epochs=20,
steps_per_epoch=n_train_samples // batch_size,
validation_data= val_gen,
validation_steps=n_val_samples // batch_size)
As a result of this code, we'll be able to see that a simple RNN does much better
than the convolutional model, with a loss of around 1,548,653. You'll remember that
previously our loss was at 12,793,928. However, we can do much better using a more
sophisticated version of the RNN.
LSTM
In the last section, we learned about basic RNNs. In theory, simple RNNs should be
able to retain even long-term memories. However, in practice, this approach often
falls short because of the vanishing gradients problem.
Over the course of many timesteps, the network has a hard time keeping up
meaningful gradients. While this is not the focus of this chapter, a more detailed
exploration of why this happens can be read in the 1994 paper, Learning long-term
dependencies with gradient descent is difficult, available at -https://ieeexplore.ieee.
org/document/279181 - by Yoshua Bengio, Patrice Simard, and Paolo Frasconi.
In direct response to the vanishing gradients problem of simple RNNs, the Long
Short-Term Memory (LSTM) layer was invented. This layer performs much better
at longer time series. Yet, if relevant observations are a few hundred steps behind in
the series, then even LSTM will struggle. This is why we manually included some
lagged observations.
[ 157 ]
Understanding Time Series
Before we dive into details, let's look at a simple RNN that has been unrolled
over time:
As you can see, this is the same as the RNN that we saw in Chapter 2, Applying
Machine Learning to Structured Data, except that this has been unrolled over time.
The carry
The central addition of an LSTM over an RNN is the carry. The carry is like
a conveyor belt that runs along the RNN layer. At each time step, the carry is fed
into the RNN layer. The new carry gets computed from the input, RNN output, and
old carry, in a separate operation from the RNN layer itself::
[ 158 ]
Chapter 4
To understand what the Compute Carry is, we should determine what should be
added from the input and state:
In these formulas st is the state at time t (output of the simple RNN layer), int is the
input at time t, and Ui, Wi, Uk, and Wk are the model parameters (matrices) that will
be learned. a() is an activation function.
To determine what should be forgotten from the state and input, we need to use the
following formula:
ct +1 = ct ∗ ft + it ∗ kt
While the standard theory claims that the LSTM layer learns what to add and what
to forget, in practice, nobody knows what really happens inside an LSTM. However,
LSTM models have been shown to be quite effective at learning long-term memory.
Take this time to note that LSTM layers do not need an extra activation function as
they already come with a tanh activation function out of the box.
LSTMs can be used in the same way as SimpleRNN:
from keras.layers import LSTM
model = Sequential()
model.add(LSTM(16,input_shape=(max_len,n_features)))
model.add(Dense(1))
To stack layers, you also need to set return_sequences to True. Note that you can
easily combine LSTM and SimpleRNN using the following code:
model = Sequential()
model.add(LSTM(32,return_sequences=True,
input_shape=(max_len,n_features)))
model.add(SimpleRNN(16, return_sequences = True))
model.add(LSTM(16))
model.add(Dense(1))
[ 159 ]
Understanding Time Series
Note: If you are using a GPU and TensorFlow backend with Keras, use
CuDNNLSTM instead of LSTM. It's significantly faster while working in
exactly the same way.
We'll now compile and run the model just as we did before:
model.compile(optimizer='adam',
loss='mean_absolute_percentage_error')
model.fit_generator(train_gen,
epochs=20,
steps_per_epoch=n_train_samples // batch_size,
validation_data= val_gen,
validation_steps=n_val_samples // batch_size)
This time, the loss went as low as 88,735, which is several orders of magnitude better
than our initial model.
Recurrent dropout
Having read this far into the book, you've already encountered the concept of
dropout. Dropout removes some elements of one layer of input at random. A common
and important tool in RNNs is a recurrent dropout, which does not remove any inputs
between layers but inputs between time steps:
[ 160 ]
Chapter 4
Just as with regular dropout, recurrent dropout has a regularizing effect and
can prevent overfitting. It's used in Keras by simply passing an argument to the
LSTM or RNN layer.
As we can see in the following code, recurrent dropout, unlike regular dropout,
does not have its own layer:
model = Sequential()
model.add(LSTM(16,
recurrent_dropout=0.1,
return_sequences=True,
input_shape=(max_len,n_features)))
model.add(LSTM(16,recurrent_dropout=0.1))
model.add(Dense(1))
The key idea in Bayesian deep learning is that there is inherent uncertainty in
the model. Sometimes this is done by learning a mean and standard deviation for
weights instead of just a single weight value. However, this approach increases the
number of parameters required, so it did not catch on. A simpler hack that allows
us to turn regular deep networks into Bayesian deep networks is to activate dropout
during prediction time and then make multiple predictions.
In this section, we will be using a simpler dataset than before. Our X values are
20 random values between -5 and 5, and our y values are just the sine function
applied to these values.
[ 161 ]
Understanding Time Series
Our neural network is relatively straightforward, too. Note that Keras does not allow
us to make a dropout layer the first layer, therefore we need to add a Dense layer
that just passes through the input value. We can achieve this with the following code:
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation
model = Sequential()
model.add(Dense(1,input_dim = 1))
model.add(Dropout(0.05))
model.add(Dense(20))
model.add(Activation('relu'))
model.add(Dropout(0.05))
model.add(Dense(20))
model.add(Activation('relu'))
model.add(Dropout(0.05))
model.add(Dense(20))
model.add(Activation('sigmoid'))
model.add(Dense(1))
To fit this function, we need a relatively low learning rate, so we import the Keras
vanilla stochastic gradient descent optimizer in order to set the learning rate there.
We then train the model for 10,000 epochs. Since we are not interested in the training
logs, we set verbose to 0, which makes the model train "quietly."
We want to test our model over a larger range of values, so we create a test dataset
with 200 values ranging from -10 to 10 in 0.1 intervals. We can imitate the test by
running the following code:
X_test = np.arange(-10,10,0.1)
X_test = np.expand_dims(X_test,-1)
And now comes the magic trick! Using keras.backend, we can pass settings to
TensorFlow, which runs the operations in the background. We use the backend to
set the learning phase parameter to 1. This makes TensorFlow believe that we are
training, and so it will apply dropout. We then make 100 predictions for our test
data. The result of these 100 predictions is a probability distribution for the y value
at every instance of X.
[ 162 ]
Chapter 4
Note: For this example to work, you have to load the backend, clear
the session, and set the learning phase before defining and training the
model, as the training process will leave the setting in the TensorFlow
graph. You can also save the trained model, clear the session, and reload
the model. See the code for this section for a working implementation.
And now we can obtain our distributions with the following code:
probs = []
for i in range(100):
out = model.predict(X_test)
probs.append(out)
Next we can calculate the mean and standard deviation for our distributions:
p = np.array(probs)
mean = p.mean(axis=0)
std = p.std(axis=0)
Finally, we plot the model's predictions with one, two, and four standard deviations
(corresponding to different shades of blue):
plt.figure(figsize=(10,7))
plt.plot(X_test,mean,c='blue')
plt.scatter(X,y,c='black')
[ 163 ]
Understanding Time Series
As you can see, the model is relatively confident around areas where it had data
and becomes less and less confident the further away it gets from the data points.
Getting uncertainty estimates from our model increases the value we can get from
it. It also helps in improving the model if we can detect where the model is over or
under confident. Right now, Bayesian deep learning is only in its infancy, and we
will certainly see many advances in the next few years.
Exercises
Now we're at the end of the chapter, why not try some of the following exercises?
You'll find guides on how to complete them all throughout this chapter:
[ 164 ]
Chapter 4
Summary
In this chapter, you learned about a wide range of conventional tools for dealing
with time series data. You also learned about one-dimensional convolution and
recurrent architectures, and finally, you learned a simple way to get your models
to express uncertainty.
Time series are the most iconic form of financial data. This chapter has given you
a rich toolbox for dealing with time series. Let's recap all of the things that we've
covered on the example of forecasting web traffic for Wikipedia:
[ 165 ]
Understanding Time Series
This rich toolbox of time series techniques comes in especially handy in the next
chapter, where we will cover natural language processing. Language is basically
a sequence, or time series, of words. This means we can reuse many tools from time
series modeling for natural language processing.
In the next chapter, you will learn how to find company names in text, how to group
text by topic, and even how to translate text using neural networks.
[ 166 ]
Parsing Textual Data with
Natural Language Processing
It's no accident that Peter Brown, Co-CEO of Renaissance Technologies, one of
the most successful quantitative hedge funds of all time, had previously worked
at IBM, where he applied machine learning to natural language problems.
Natural language processing (NLP) is making inroads into the finance sector.
As an example, insurance companies are increasingly looking to process claims
automatically, while retail banks try to streamline their customer service and offer
better products to their clients. The understanding of text is increasingly becoming
the go-to application of machine learning within the finance sector.
Historically, NLP has relied on hand-crafted rules that were created by linguists.
Today, the linguists are being replaced by neural networks that are able to learn
the complex, and often hard to codify, rules of language.
In this chapter, you will learn how to build powerful natural language models with
Keras, as well as how to use the spaCy NLP library.
[ 167 ]
Parsing Textual Data with Natural Language Processing
To use spaCy locally, you will need to install the library and download its pretrained
models separately.
Note: This chapter makes use of the English language models, but more
are available. Most features are available in English, German, Spanish,
Portuguese, French, Italian, and Dutch. Entity recognition is available
for many more languages through the multi-language model.
The core of spaCy is made up of the Doc and Vocab classes. A Doc instance
contains one document, including its text, tokenized version, and recognized
entities. The Vocab class, meanwhile, keeps track of all the common information
found across documents.
[ 168 ]
Chapter 5
spaCy is useful for its pipeline features, which contain many of the parts needed for
NLP. If this all seems a bit abstract right now, don't worry, as this section will show
you how to use spaCy for a wide range of practical tasks.
You can find the data and code for this section on Kaggle at https://
www.kaggle.com/jannesklaas/analyzing-the-news.
The data that we'll use for this first section is from a collection of 143,000 articles
taken from 15 American publications. The data is spread out over three files. We
are going to load them separately, merge them into one large DataFrame, and then
delete the individual DataFrames in order to save memory.
df = pd.concat([a1,a2,a3])
As a result of running the preceding code, the data will end up looking like this:
After getting our data to this state, we can then plot the distribution of publishers to
get an idea of what kind of news we are dealing with.
[ 169 ]
Parsing Textual Data with Natural Language Processing
After successfully running this code, we'll see this chart showing the distribution of
news sources from our dataset:
As you can see in the preceding graph the dataset that we extracted contains no
articles from classical financial news media, instead it mostly contains articles from
mainstream and politically oriented publications.
The first thing we need to do is load spaCy, in addition to the model for English
language processing:
import spacy
nlp = spacy.load('en')
[ 170 ]
Chapter 5
Next, we must select the text of the article from our data:
text = df.loc[0,'content']
Finally, we'll run this piece of text through the English language model pipeline.
This will create a Doc instance, something we explained earlier on in this chapter.
The file will hold a lot of information, including the named entities:
doc = nlp(text)
One of the best features of spaCy is that it comes with a handy visualizer called
displacy, which we can use to show the named entities in text. To get the visualizer
to generate the display, based on the text from our article, we must run this code:
from spacy import displacy
displacy.render(doc, #1
style='ent', #2
jupyter=True) #3
With that command now executed, we've done three important things, which are:
[ 171 ]
Parsing Textual Data with Natural Language Processing
And voilà! As you can see, there are a few mishaps, such as blank spaces being
classified as organizations, and "Obama" being classified as a place.
So, why has this happened? It's because the tagging has been done by a neural
network and neural networks are strongly dependent on the data that they were
trained on. So, because of these imperfections, we might find that we need to
fine-tune the tagging model for our own purposes, and in a minute, we will
see how that works.
You can also see in our output that NER offers a wide range of tags, some of which
come with strange abbreviations. For now, don't worry as we will examine a full list
of tags later on in this chapter.
Right now, let's answer a different question: what organizations does the news in our
dataset write about? To make this exercise run faster, we will create a new pipeline
in which we will disable everything but NER.
To find out the answer to this question, we must first run the following code:
nlp = spacy.load('en',
disable=['parser',
'tagger',
'textcat'])
In the next step, we'll loop over the first 1,000 articles from our dataset, which can be
done with the following code:
from tqdm import tqdm_notebook
frames = []
for i in tqdm_notebook(range(1000)):
doc = df.loc[i,'content'] #1
text_id = df.loc[i,'id'] #2
doc = nlp(doc) #3
ents = [(e.text, e.start_char, e.end_char, e.label_) #4
for e in doc.ents
if len(e.text.strip(' -—')) > 0]
frame = pd.DataFrame(ents) #5
frame['id'] = text_id #6
frames.append(frame) #7
npf = pd.concat(frames) #8
npf.columns = ['Text','Start','Stop','Type','id'] #9
The code we've just created has nine key points. Let's take a minute to break it down,
so we are confident in understanding what we've just written. Note that in the
preceding code, the hashtag, #, refers to the number it relates to in this following list:
Now that we've done that, the next step is to plot the distribution of the types of
entities that we found. This code will produce a chart which can be created with
the following code:
npf.Type.value_counts().plot(kind='bar')
[ 173 ]
Parsing Textual Data with Natural Language Processing
After seeing the preceding graph, it is a fair question to ask which categories spaCy
can identify and where they come from. The English language NER that comes
with spaCy is a neural network trained on the OntoNotes 5.0 corpus, meaning it can
recognize the following categories:
Using this list, we will now look at the 15 most frequently named organizations,
categorized as ORG. As part of this, we will produce a similar graph showing us that
information.
[ 174 ]
Chapter 5
As you can see, political institutions such as the senate are most frequently named
in our news dataset. Likewise, some companies, such as Volkswagen, that were in the
center of media attention can also be found in the chart. Take a minute to also notice
how the White House and White House are listed as two separate organizations,
despite us knowing they are the same entity.
Should you want to look at it in more detail, you can get the documentation and
example from the following link: https://pandas.pydata.org/pandas-docs/
stable/generated/pandas.Series.str.replace.html
[ 175 ]
Parsing Textual Data with Natural Language Processing
Also, note how Trump is shown here as an organization. However, if you look at
the tagged text, you will also see that "Trump" is tagged several times as an NORP,
a political organization. This has happened because the NER infers the type of tag
from the context. Since Trump is the U.S. president, his name often gets used in the
same context as (political) organizations.
This pretrained NER gives you a powerful tool that can solve many common NLP
tasks. So, in reality, from here you could conduct all kinds of other investigations.
For example, we could fork the notebook to see whether The New York Times is
mentioned as different entities more often than the Washington Post or Breitbart.
As you can see, you provide a list of tuples of the string, together with the start
and end points, as well as the types of entities you want to tag. Data such as this
is usually collected through manual tagging, often on platforms such as Amazon's
Mechanical Turk (MTurk).
The company behind spaCy, Explosion AI, also make a (paid) data tagging system
called Prodigy, which enables efficient data collection. Once you have collected enough
data, you can either fine-tune a pretrained model or initialize a completely new model.
Alternatively, to create a new and empty model from scratch that is ready for the
English language, use the blank function:
nlp = spacy.blank('en')
[ 176 ]
Chapter 5
Either way, we need to get access to the NER component. If you have created a blank
model, you'll need to create an NER pipeline component and add it to the model.
If you have loaded an existing model, you can just access its existing NER by running
the following code:
if 'ner' not in nlp.pipe_names:
ner = nlp.create_pipe('ner')
nlp.add_pipe(ner, last=True)
else:
ner = nlp.get_pipe('ner')
The next step is to ensure that our NER can recognize the labels we have. Imagine
our data contained a new type of named entity such as ANIMAL. With the add_label
function, we can add a label type to an NER.
The code to achieve this can be seen below, but don't worry if it doesn't make sense
right now, we'll break it down on the next page:
for _, annotations in TRAIN_DATA:
for ent in annotations.get('entities'):
ner.add_label(ent[2])
import random
#1
other_pipes = [pipe for pipe in nlp.pipe_names if pipe != 'ner']
with nlp.disable_pipes(*other_pipes):
optimizer = nlp._optimizer #2
if not nlp._optimizer:
optimizer = nlp.begin_training()
for itn in range(5): #3
random.shuffle(TRAIN_DATA) #4
losses = {} #5
for text, annotations in TRAIN_DATA: #6
nlp.update( #7
[text],
[annotations],
drop=0.5, #8
sgd=optimizer, #9
losses=losses) #10
print(losses)
[ 177 ]
Parsing Textual Data with Natural Language Processing
1. We disable all pipeline components that are not the NER by first getting a
list of all the components that are not the NER and then disabling them for
training.
2. Pretrained models come with an optimizer. If you have a blank model, you
will need to create a new optimizer. Note that this also resets the model
weights.
3. We now train for a number of epochs, in this case, 5.
4. At the beginning of each epoch, we shuffle the training data using Python's
built-in random module.
5. We create an empty dictionary to keep track of the losses.
6. We then loop over the text and annotations in the training data.
7. nlp.update performs one forward and backward pass, and updates
the neural network weights. We need to supply text and annotations,
so that the function can figure out how to train a network from it.
8. We can manually specify the dropout rate we want to use while training.
9. We pass a stochastic gradient descent optimizer that performs the model
updates. Note that you cannot just pass a Keras or TensorFlow optimizer
here, as spaCy has its own optimizers.
10. We can also pass a dictionary to write losses that we can later print
to monitor progress.
Once you've run the code, the output should look something like this:
{'ner': 5.0091189558407585}
{'ner': 3.9693684224622108}
{'ner': 3.984836024903589}
{'ner': 3.457960373417813}
{'ner': 2.570318400714134}
What you are seeing is the loss value of a part of the spaCy pipeline, in this case,
the named entity recognition (NER) engine. Similar to the cross-entropy loss we
discussed in previous chapters, the actual value is hard to interpret and does not tell
you very much. What matters here is that the loss is decreasing over time and that
it reaches a value much lower than the initial loss.
The algorithmic trading systems obviously failed here to understand that such an
acquisition would be impossible as Apple had a market capitalization of $800 billion at
the time, coupled with the fact that the move would likely not find regulatory approval.
So, the question arises, why did the trading algorithms choose to buy stock based on
these four words? The answer is through part-of-speech (POS) tagging. POS tagging
allows an understanding of which words take which function in a sentence and how
the words relate to each other.
spaCy comes with a handy, pretrained POS tagger. In this section we're going to
apply this to the Google/Apple news story. To start the POS tagger, we need to run
the following code:
import spacy
from spacy import displacy
nlp = spacy.load('en')
Again, we will load the pretrained English model and run our sentence through it.
Then we'll use displacy just as we did for NER.
To make the graphics fit better in this book, we will set the distance option to
something shorter than the default, in this case, 1,120, so that words get displayed
closer together, as we can see in the following diagram:
As you can see, the POS tagger identified buy as a verb and Google and Apple
as the nouns in the sentence. It also identified that Apple is the object the action is
applied to and that Google is applying the action.
[ 179 ]
Parsing Textual Data with Natural Language Processing
After running the preceding code, we get the following table featured as the result:
In our example, Google is the root of the sentence, while Apple is the object of the
sentence. The verb applied to Apple is "buy."
Rule-based matching
Before deep learning and statistical modeling took over, NLP was all about rules.
That's not to say that rule-based systems are dead! They are often easy to set up and
perform very well when it comes to doing simple tasks.
Imagine you wanted to find all mentions of Google in a text. Would you really train
a neural network-based named entity recognizer? If you did, you would have to
run all of the text through the neural network and then look for Google in the entity
texts. Alternatively, would you rather just search for text that exactly matches Google
with a classic search algorithm? Well, we're in luck, as spaCy comes with an easy-to-
use, rule-based matcher that allows us to do just that.
Before we start this section, we first must make sure that we reload the English
language model and import the matcher. This is a very simple task that can be
done by running the following code:
import spacy
from spacy.matcher import Matcher
nlp = spacy.load('en')
[ 180 ]
Chapter 5
This pattern is fulfilled if the lower case first token is hello. The LOWER attribute
checks if both words would match if they were both converted to lowercase. That
means if the actual token text is "Hello" or "HELLO," then it would also fulfill the
requirement. The second token has to be punctuation to pick up the comma, so the
phrases "hello. world" or "hello! world" would both work, but not "hello world."
The lower case of the third token has to be "world," so "WoRlD" would also be fine.
[ 181 ]
Parsing Textual Data with Natural Language Processing
In order to add the required attributes to our matcher, we must call the following:
matcher.add('HelloWorld', None, pattern)
The add function expects three arguments. The first is a name of the pattern, in
this case, HelloWorld, so that we can keep track of the patterns we added. The
second is a function that can process matches once found. Here we pass None,
meaning no function will be applied, though we will use this tool later. Finally,
we need to pass the list of token attributes we want to search for.
To use our matcher, we can simply call matcher(doc). This will give us back
all the matches that the matcher found. We can call this by running the following:
doc = nlp(u'Hello, world! Hello world!')
matches = matcher(doc)
The first thing in a match is the hash of the string found. This is just to identify what
was found internally; we won't use it here. The next two numbers indicate the range
in which the matcher found something, here tokens 0 to 3.
Hello, world
In the next section we will look at how we can add custom functions to matchers.
Let's build a rule-based matcher that always classifies the word "iPhone" as a product
entity.
First, we have to get the hash of the word PRODUCT. Words in spaCy can be uniquely
identified by their hash. Entity types also get identified by their hash. To set an entity
of the product type, we have to be able to provide the hash for the entity name.
[ 182 ]
Chapter 5
We can get the name from the language model's vocabulary by running the following:
PRODUCT = nlp.vocab.strings['PRODUCT']
Next, we need to define an on_match rule. This function will be called every time
the matcher finds a match. on_match rules have four arguments:
1. We index all matches to find our match at index i. One match is a tuple
of a match_id, the start of the match, and the end of the match.
2. We add a new entity to the document's named entities. An entity is a tuple of
the hash of the type of entity (the hash of the word PRODUCT here), the start of
the entity, and the end of the entity. To append an entity, we have to nest it
in another tuple. Tuples that only contain one value need to include a comma
at the end. It is important not to overwrite doc.ents, as we otherwise would
remove all the entities that we have already found.
We should note that matchers allow us to add multiple patterns, so we can add
a matcher for just the word "iPhone" and another pattern for the word "iPhone"
together with a version number, as in "iPhone 5":
pattern1 = [{'LOWER': 'iPhone'}] #1
pattern2 = [{'ORTH': 'iPhone'}, {'IS_DIGIT': True}] #2
matcher = Matcher(nlp.vocab) #3
matcher.add('iPhone', add_product_ent,pattern1, pattern2) #4
[ 183 ]
Parsing Textual Data with Natural Language Processing
We will now pass one of the news articles through the matcher:
doc = nlp(df.content.iloc[14]) #1
matches = matcher(doc) #2
Now that the matcher is set up, we need to add it to the pipeline so that spaCy can
use it automatically. This will be the focus in the next section.
The spaCy pipeline calls the components of the pipeline as functions and always
expects the annotated document to be returned. Returning anything else could
break the pipeline.
We can then add the matcher to the main pipeline, as can be seen in the
following code:
nlp.add_pipe(matcher_component,last=True)
The matcher is now the last piece of the pipeline. From this point onward iPhones
will now get tagged based on the matcher's rules.
And boom! All mentions of the word "iPhone" (case independent) are now tagged
as named entities of the product type. You can validate this by displaying the entities
with displacy as we have done in the following code:
displacy.render(doc,style='ent',jupyter=True)
[ 184 ]
Chapter 5
A matcher looking for this pattern would pick up any combination of a person's
name, for example, John Appleseed, or Daniel; any version of the word receive,
for example, received, receives, and so on; followed by an expression of money,
for example, $4 million.
This matcher could be run over a large text corpus with the on_match rule handily
saving the found snippets into a database. The machine learning approach for
naming entities and the rule-based approach go hand in hand seamlessly.
Since there is much more training data available with annotations for names
and money, rather than statements about executive education, it is much easier
to combine the NER with a rule-based method rather than training a new NER.
Regular expressions
Regular expressions, or regexes, are a powerful form of rule-based matching.
Invented back in the 1950s, they were, for a very long time, the most useful way
to find things in text and proponents argue that they still are.
No chapter on NLP would be complete without mentioning regexes. With that being
said, this section is by no means a complete regex tutorial. It's intended to introduce
the general idea and show how regexes can be used in Python, pandas, and spaCy.
[ 185 ]
Parsing Textual Data with Natural Language Processing
A very simple regex pattern could be "a." This would only find instances of the
lower-case letter a followed by a dot. However, regexes also allow you to add
ranges of patterns; for example, "[a-z]." would find any lower-case letter followed
by a dot, and "xy." would find only the letters "x" or "y" followed by a dot.
Regex patterns are case sensitive, so "A-Z" would only capture upper-case letters.
This is useful if we are searching for expressions in which the spelling is frequently
different; for example, the pattern "seriali[sz]e" would catch the British as well as the
American English version of the word.
The same goes for numbers. "0-9" captures all numbers from 0 to 9. To find
repetitions, you can use "*," which captures zero or more occurrences, or "+," which
captures one or more occurrences. For example, "[0-9]+" would capture any series of
numbers, which might be useful when looking for years. While "[A-Z][a-z] + [0-9] +,"
for example, would find all words starting with a capital letter followed by one or
more digit, such as "March 2018" but also "Jaws 2."
Curly brackets can be used to define the number of repetitions. For instance, "[0-9]
{4}" would find number sequences with exactly four digits. As you can see, a regex
does not make any attempt to understand what is in the text, but rather offers
a clever method of finding text that matches patterns.
A practical use case in the financial industry is finding the VAT number of
companies in invoices. These follow a pretty strict pattern in most countries that
can easily be encoded. VAT numbers in the Netherlands, for example, follow this
regex pattern: "NL[0-9]{9}B[0-9]{2}".
[ 186 ]
Chapter 5
So, to find all the occurrences of a BTW number in the string, we can call
re.findall, which will return a list of all strings matching the pattern found.
To call this, we simply run:
re.findall(pattern,my_string)
['NL003028112B01']
re also allows the passing of flags to make the development of regex patterns a bit
easier. For example, to ignore the case of letters when matching a regular expression,
we can add a re.IGNORECASE flag, like we've done here:
re.findall(pattern,my_string, flags=re.IGNORECASE)
Often, we are interested in a bit more information about our matches. To this end,
there is a match object. re.search yields a match object for the first match found:
match = re.search(pattern,my_string)
We can get more information out of this object, such as the location of our match,
simply by running:
match.span()
(18, 32)
The span, the start and the end of our match, is the characters 18 to 32.
Regex in pandas
The data for NLP problems often comes in pandas DataFrames. Luckily for us,
pandas natively supports regex. If, for example, we want to find out whether any of
the articles in our news dataset contain a Dutch BTW number, then we can pass the
following code:
df[df.content.str.contains(pattern)]
This would yield all the articles that include a Dutch BTW number, but
unsurprisingly no articles in our dataset do.
[ 187 ]
Parsing Textual Data with Natural Language Processing
Note: In preparation for this section, the code and data can be found on
Kaggle at https://www.kaggle.com/jannesklaas/nlp-disasters.
Sanitizing characters
To store text, computers need to encode the characters into bits. There are several
different ways to do this, and not all of them can deal with all the characters out there.
It is good practice to keep all the text files in one encoding scheme, usually UTF-8,
but of course, that does not always happen. Files might also be corrupted, meaning
that a few bits are off, therefore rendering some characters unreadable. Therefore,
before we do anything else, we need to sanitize our inputs.
Python offers a helpful codecs library, which allows us to deal with different
encodings. Our data is UTF-8 encoded, but there are a few special characters in
there that cannot be read easily. Therefore, we have to sanitize our text of these
special characters, which we can do by running the following:
import codecs
input_file = codecs.open('../input/socialmedia-disaster-tweets-
DFE.csv',
'r',',
encoding='utf-8',
errors='replace')
Now that's done, all we have to do is loop over the lines in our input file that we read
with our codecs reader and save it as a regular CSV file again. We can achieve this
by running the following:
for line in input_file:
out = line
output_file.write(line)
Likewise, it's good practice to close the file objects afterward, which we can do by
running:
input_file.close()
output_file.close()
[ 189 ]
Parsing Textual Data with Natural Language Processing
Lemmatization
Lemmas have already made several appearances throughout this chapter. A lemma
in the field of linguistics, also called a headword, is the word under which the set of
related words or forms appears in a dictionary. For example, "was" and "is" appear
under "be," "mice" appears under "mouse," and so on. Quite often, the specific form
of a word does not matter very much, so it can be a good idea to convert all your text
into its lemma form.
spaCy offers a handy way to lemmatize text, so once again, we're going to load
a spaCy pipeline. Only that in this case, we don't need any pipeline module aside
from the tokenizer. The tokenizer splits the text into separate words, usually by
spaces. These individual words, or tokens, can then be used to look up their lemma.
In our case, it looks like this:
import spacy
nlp = spacy.load('en',disable=['tagger','parser','ner'])
Lemmatization can be slow, especially for big files, so it makes sense to track our
progress. tqdm allows us to show progress bars on the pandas apply function. All we
have to do is import tqdm as well as the notebook component for pretty rendering in
our work environment. We then have to tell tqdm that we would like to use it with
pandas. We can do this by running the following:
from tqdm import tqdm, tqdm_notebook
tqdm.pandas(tqdm_notebook)
We can now run progress_apply on a DataFrame just as we would use the standard
apply method, but here it has a progress bar.
For each row, we loop over the words in the text column and save the lemma of the
word in a new lemmas column:
df['lemmas'] = df["text"].progress_apply(lambda row:
[w.lemma_ for w in nlp(row)])
Our lemmas column is now full of lists, so to turn the lists back into text, we will
join all of the elements of the lists with a space as a separator, as we can see in the
following code:
df['joint_lemmas'] = df['lemmas'].progress_apply
(lambda row: ' '.join(row))
[ 190 ]
Chapter 5
The tweets where humans cannot decide whether it is about a real disaster are not
interesting to us. Therefore, we will just remove the category, Can't Decide, which we
can do in the following code:
df = df[df.choose_one != "Can't Decide"]
We are also only interested in mapping text to relevance, therefore we can drop all
the other metadata and just keep these two columns, which we do here:
df = df[['text','choose_one']]
Finally, we're going to convert the target into numbers. This is a binary classification
task, as there are only two categories. So, we map Relevant to 1 and Not Relevant to 0:
f['relevant'] = df.choose_one.map({'Relevant':1,'Not Relevant':0})
Bag-of-words
A simple yet effective way of classifying text is to see the text as a bag-of-words. This
means that we do not care for the order in which words appear in the text, instead
we only care about which words appear in the text.
Picture this as an example: the count vector for the text "I see cats and dogs and
elephants" could look like this:
In reality, count vectors are pretty sparse. There are about 23,000 different words
in our text corpus, so it makes sense to limit the number of words we want to include
in our count vectors. This could mean excluding words that are often just gibberish
or typos with no meaning. As a side note, if we kept all the rare words, this could
be a source of overfitting.
Our count vectorizer can now transform texts into count vectors. Each count vector
will have 10,000 dimensions:
X_train_counts = count_vectorizer.fit_transform(X_train)
X_test_counts = count_vectorizer.transform(X_test)
Once we have obtained our count vectors, we can then perform a simple logistic
regression on them. While we could use Keras for logistic regression, as we did in the
first chapter of this book, it is often easier to just use the logistic regression class from
scikit-learn:
from sklearn.linear_model import LogisticRegression
clf = LogisticRegression()
clf.fit(X_train_counts, y_train)
y_predicted = clf.predict(X_test_counts)
Now that we have predictions from our logistic regressor, we can measure the
accuracy of it with sklearn:
from sklearn.metrics import accuracy_score
accuracy_score(y_test, y_predicted)
0.8011049723756906
[ 192 ]
Chapter 5
As you can see, we've got 80% accuracy, which is pretty decent for such a simple
method. A simple count vector-based classification is useful as a baseline for more
advanced methods, which we will be discussing later.
TF-IDF
TF-IDF stands for Term Frequency, Inverse Document Frequency. It aims
to address a problem of simple word counting, that being words that frequently
appear in a text are important, while words that appear in all texts are not important.
The TF component is just like a count vector, except that TF divides the counts by
the total number of words in a text. Meanwhile, the IDF component is the logarithm
of the total number of texts in the entire corpus divided by the number of texts that
include a specific word.
TF-IDF is the product of these two measurements. TF-IDF vectors are like count
vectors, except they contain the TF-IDF scores instead of the counts. Rare words
will gain a high score in the TF-IDF vector.
X_train_tfidf = tfidf_vectorizer.fit_transform(X_train)
X_test_tfidf = tfidf_vectorizer.transform(X_test)
Once we have the TF-IDF vectors, we can train a logistic regressor on them just like
we did for count vectors:
clf_tfidf = LogisticRegression()
clf_tfidf.fit(X_train_tfidf, y_train)
y_predicted = clf_tfidf.predict(X_test_tfidf)
In this case, TF-IDF does slightly worse than count vectors. However, because the
performance difference is very small, this poorer performance might be attributable
to chance in this case:
accuracy_score(y_pred=y_predicted, y_true=y_test)
0.7978821362799263
[ 193 ]
Parsing Textual Data with Natural Language Processing
Topic modeling
A final, very useful application of word counting is topic modeling. Given a set
of texts, are we able to find clusters of topics? The method to do this is called
Latent Dirichlet Allocation (LDA).
Note: The code and data for this section can be found on Kaggle at
https://www.kaggle.com/jannesklaas/topic-modeling-
with-lda.
While the name is quite a mouth full, the algorithm is a very useful one, so we will
look at it step by step. LDA makes the following assumption about how texts are
written:
1. First, a topic distribution is chosen, say 70% machine learning and 30%
finance.
2. Second, the distribution of words for each topic is chosen. For example,
the topic "machine learning" might be made up of 20% the word "tensor,"
10% the word "gradient," and so on. This means that our topic distribution
is a distribution of distributions, also called a Dirichlet distribution.
3. Once the text gets written, two probabilistic decisions are made for each
word: first, a topic is chosen from the distribution of topics in the document.
Then, a word is chosen for the distribution of words in that document.
Note that not all documents in a corpus have the same distribution of topics. We
need to specify a fixed number of topics. In the learning process, we start out by
assigning each word in the corpus randomly to one topic. For each document,
we then calculate the following:
p (t | d )
p (w | t )
That is the probability of a word, w, to belong to a topic, t. We then assign the word
to a new topic, t, with the following probability:
p (t | d ) ∗ p ( w | t )
[ 194 ]
Chapter 5
In other words, we assume that all of the words are already correctly assigned to a
topic except for the word currently under consideration. We then try to assign words
to topics to make documents more homogenous in their topic distribution. This way,
words that actually belong to a topic cluster together.
Scikit-learn offers an easy-to-use LDA tool that will help us achieve this. To use
this, we must first create a new LDA analyzer and specify the number of topics,
called components that we expect.
We then create count vectors, just as we did for the bag-of-words analysis. For
LDA, it is important to remove frequent words that don't mean anything, such
as "an" or "the," so-called stop words. CountVectorizer comes with a built-in
stopword dictionary that removes these words automatically. To use this, we'll need
to run the following code:
from sklearn.feature_extraction.text import TfidfVectorizer,
CountVectorizer
vectorizer = CountVectorizer(stop_words='english')
tf = vectorizer.fit_transform(df['joint_lemmas'])
To inspect our results, we can print out the most frequent words for each topic. To this
end, we first need to specify the number of words per topic we want to print, in this
case 5. We also need to extract the mapping word count vector indices to words:
n_top_words = 5
tf_feature_names = vectorizer.get_feature_names()
Now we can loop over the topics of the LDA, in order to print the most frequent
words:
for topic_idx, topic in enumerate(lda.components_):
message = "Topic #%d: " % topic_idx
message += " ".join([tf_feature_names[i]
for i in topic.argsort()[:-
n_top_words - 1:-1]])
print(message)
Topic #0: http news bomb kill disaster
Topic #1: pron http like just https
[ 195 ]
Parsing Textual Data with Natural Language Processing
As you can see, the LDA seems to have discovered the grouping into serious tweets
and non-serious ones by itself without being given the targets.
This method is very useful for classifying news articles, too. Back in the world of
finance, investors might want to know if there is a news article mentioning a risk
factor they are exposed to. The same goes for support requests for consumer-facing
organizations, which can be clustered this way.
Word embeddings
The order of words in a text matters. Therefore, we can expect higher performance
if we do not just look at texts in aggregate but see them as a sequence. This section
makes use of a lot of the techniques discussed in the previous chapter; however,
here we're going to add a critical ingredient, word vectors.
Words and word tokens are categorical features. As such, we cannot directly feed
them into a neural network. Previously, we have dealt with categorical data by
turning it into one-hot encoded vectors. Yet for words, this is impractical. Since our
vocabulary is 10,000 words, each vector would contain 10,000 numbers that are all
zeros except for one. This is highly inefficient, so instead, we will use an embedding.
In practice, embeddings work like a lookup table. For each token, they store a vector.
When the token is given to the embedding layer, it returns the vector for that token
and passes it through the neural network. As the network trains, the embeddings get
optimized as well.
Remember that neural networks work by calculating the derivative of the
loss function with respect to the parameters (weights) of the model. Through
backpropagation, we can also calculate the derivative of the loss function with
respect to the input of the model. Thus we can optimize the embeddings to deliver
ideal inputs that help our model.
The Tokenizer class also allows us to specify how many words we want to consider,
so once again we will only use the 10,000 most used words, which we can specify
by running:
from keras.preprocessing.text import Tokenizer
import numpy as np
max_words = 10000
The tokenizer works a lot like CountVectorizer from sklearn. First, we create
a new tokenizer object. Then we fit the tokenizer, and finally, we can transform the
text into tokenized sequences:
tokenizer = Tokenizer(num_words=max_words)
tokenizer.fit_on_texts(df['joint_lemmas'])
sequences = tokenizer.texts_to_sequences(df['joint_lemmas'])
The sequences variable now holds all of our texts as numeric tokens. We can
look up the mapping of words to numbers from the tokenizer's word index with
the following code:
word_index = tokenizer.word_index
print('Token for "the"',word_index['the'])
print('Token for "Movie"',word_index['movie'])
Token for "the" 4
Token for "Movie" 333
As you can see, frequently used words such as "the" have lower token numbers
than less frequent words such as "movie." You can also see that word_index is a
dictionary. If you are using your model in production, you can save this dictionary
to disk in order to convert words into tokens at a later time.
Finally, we need to turn our sequences into sequences of equal length. This is not
always necessary, as some model types can deal with sequences of different lengths,
but it usually makes sense and is often required. We will examine which models
need equal length sequences in the next section on building custom NLP models.
Keras' pad_sequences function allows us to easily bring all of the sequences to the
same length by either cutting off sequences or adding zeros at the end. We will bring
all the tweets to a length of 140 characters, which for a long time was the maximum
length tweets could have:
from keras.preprocessing.sequence import pad_sequences
maxlen = 140
[ 197 ]
Parsing Textual Data with Natural Language Processing
Embeddings are their own layer type in Keras. To use them, we have to specify
how large we want the word vectors to be. The 50-dimensional vector that we have
chosen to use is able to capture good embeddings even for quite large vocabularies.
Additionally, we also have to specify how many words we want embeddings for and
how long our sequences are. Our model is now a simple logistic regressor that trains
its own embeddings:
from keras.models import Sequential
from keras.layers import Embedding, Flatten, Dense
embedding_dim = 50
model = Sequential()
model.add(Embedding(max_words, embedding_dim,
input_length=maxlen))
model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))
Notice how we do not have to specify an input shape. Even specifying the input
length is only necessary if the following layers require knowledge of the input
length. Dense layers require knowledge about the input size, but since we are
using dense layers directly, we need to specify the input length here.
Word embeddings have many parameters. This is something you can see if you are
printing out the models summary:
model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
embedding_2 (Embedding) (None, 140, 50) 500000
_________________________________________________________________
flatten_2 (Flatten) (None, 7000) 0
_________________________________________________________________
dense_3 (Dense) (None, 1) 7001
=================================================================
[ 198 ]
Chapter 5
As you can see, the embedding layer has 50 parameters for 10,000 words equaling
500,000 parameters in total. This makes training slower and can increase the chance
of overfitting.
The next step is for us to compile and train our model as usual:
model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['acc'])
This model achieves about 76% accuracy on the test set but over 90% accuracy on
the training set. However, the large number of parameters in the custom embeddings
has led us to overfitting. To avoid overfitting and reduce training time, it's often
better to use pretrained word embeddings.
In this section, we will be using word embeddings trained on Wikipedia texts as well
as the Gigaword dataset. In total, the vectors were trained on a text of 6 billion tokens.
With all that being said, there are alternatives to GloVe, such as Word2Vec. Both
GloVe and Word2Vec are relatively similar, although the training method for them
is different. They each have their strengths and weaknesses, and in practice it is often
worth trying out both.
A nice feature of GloVe vectors is that they encode word meanings in vector space
so that "word algebra" becomes possible. The vector for "king" minus the vector for
"man" plus the vector for "woman," for example, results in a vector pretty close to
"queen." This means the differences between the vectors for "man" and "woman"
are the same as the differences for the vectors of "king" and "queen," as the
differentiating features for both are nearly the same.
[ 199 ]
Parsing Textual Data with Natural Language Processing
Equally, words describing similar things such as "frog" and "toad" are very close
to each other in the GloVe vector space. Encoding semantic meanings in vectors offer
a range of other exciting opportunities for document similarity and topic modeling,
as we will see later in this chapter. Semantic vectors are also pretty useful for a wide
range of NLP tasks, such as our text classification problem.
The actual GloVe vectors are in a text file. We will use the 50-dimensional
embeddings trained on 6 billion tokens. To do this, we need to open the file:
import os
glove_dir = '../input/glove6b50d'
f = open(os.path.join(glove_dir, 'glove.6B.50d.txt'))
Then we create an empty dictionary that will later map words to embeddings:
embeddings_index = {}
In the dataset, each line represents a new word embedding. The line starts with the
word, and the embedding values follow. We can read out the embeddings like this:
for line in f: #1
values = line.split() #2
word = values[0] #3
embedding = np.asarray(values[1:], dtype='float32') #4
embeddings_index[word] = embedding dictionary #5
f.close() #6
But what does that mean? Let's take a minute to break down the meaning behind the
code, which has six key elements:
1. We loop over all lines in the file. Each line contains a word and embedding.
2. We split the line by whitespace.
3. The first thing in the line is always the word.
4. Then come the embedding values. We immediately transform them into
a NumPy array and make sure that they are all floating-point numbers,
that is, decimals.
5. We then save the embedding vector in our embedding dictionary.
6. Once we are done with it, we close the file.
As a result of running this code, we now have a dictionary mapping words to their
embeddings:
print('Found %s word vectors.' % len(embeddings_index))
[ 200 ]
Chapter 5
This version of GloVe has vectors for 400,000 words, which should be enough to
cover most of the words that we will encounter. However, there might be some
words where we still do not have a vector. For these words, we will just create
random vectors. To make sure these vectors are not too far off, it is a good idea
to use the same mean and standard deviation for the random vectors as from
the trained vectors.
To this end, we need to calculate the mean and standard deviation for the GloVe
vectors:
all_embs = np.stack(embeddings_index.values())
emb_mean = all_embs.mean()
emb_std = all_embs.std()
Our embedding layer will be a matrix with a row for each word and a column for
each element of the embedding. Therefore, we need to specify how many dimensions
one embedding has. The version of GloVe we loaded earlier has 50-dimensional
vectors:
embedding_dim = 50
Next, we need to find out how many words we actually have. Although we have
set the maximum to 10,000, there might be fewer words in our corpus. At this point,
we also retrieve the word index from the tokenizer, which we will use later:
word_index = tokenizer.word_index
nb_words = min(max_words, len(word_index))
To create our embedding matrix, we first create a random matrix with the same mean
and std as the embeddings:
embedding_matrix = np.random.normal(emb_mean,
emb_std,
(nb_words, embedding_dim))
[ 201 ]
Parsing Textual Data with Natural Language Processing
This command has four key elements that we should explore in more detail before
we move on:
To use the pretrained embeddings, we just have to set the weights in the embedding
layer to the embedding matrix that we just created. To make sure the carefully
created weights are not destroyed, we are going to set the layer to be non-trainable,
which we can achieve by running the following:
model = Sequential()
model.add(Embedding(max_words,
embedding_dim,
input_length=maxlen,
weights = [embedding_matrix],
trainable = False))
model.add(Flatten())
model.add(Dense(1, activation='sigmoid'))
This model can be compiled and trained just like any other Keras model. You
will notice that it trains much faster than the model in which we trained our own
embeddings and suffers less from overfitting. However, the overall performance
on the test set is roughly the same.
Word embeddings are pretty cool in reducing training time and helping to build
accurate models. However, semantic embeddings go further. They can, for example,
be used to measure how similar two texts are on a semantical level, even if they
include different words.
Let's start with an LSTM, otherwise known as long short-term memory. All you have
to change from the implementation in the last chapter is that the first layer of the
network should be an embedding layer. This example below uses a CuDNNLSTM layer,
which trains much faster than a regular LSTM layer.
[ 202 ]
Chapter 5
Other than this, the layer remains the same. If you do not have a GPU, replace
CuDNNLSTM with LSTM:
One technique used frequently in NLP but less frequently in time series forecasting is
a bidirectional recurrent neural network (RNN). A bidirectional RNN is effectively
just two RNNs where one gets fed the sequence forward, while the other one gets fed
the sequence backward:
A bidirectional RNN
In Keras, there is a Bidirectional layer that we can wrap any RNN layer around,
such as an LSTM. We achieve this in the following code:
from keras.layers import Bidirectional
model = Sequential()
model.add(Embedding(max_words,
embedding_dim,
input_length=maxlen,
weights = [embedding_matrix], trainable =
False))
model.add(Bidirectional(CuDNNLSTM(32)))
model.add(Dense(1, activation='sigmoid'))
[ 203 ]
Parsing Textual Data with Natural Language Processing
Word embeddings are great because they enrich neural networks. They are a space-
efficient and powerful method that allows us to transform words into numbers that
a neural network can work with. With that being said, there are more advantages to
encoding semantics as vectors, such as how we can perform vector math on them!
This is useful if we want to measure the similarity between two texts, for instance.
spaCy has a built-in function to measure the similarity between two sentences.
It also comes with pretrained vectors from the Word2Vec model, which is similar
to GloVe. This method works by averaging the embedding vectors of all the words
in a text and then measuring the cosine of the angle between the average vectors.
Two vectors pointing in roughly the same direction will have a high similarity score,
whereas vectors pointing in different directions will have a low similarity score.
This is visualized in the following graph:
Similarity vectors
We can see the similarity between two phrases by running the following command:
sup1 = nlp('I would like to open a new checking account')
sup2 = nlp('How do I open a checking account?')
[ 204 ]
Chapter 5
As you can see, these requests are pretty similar, achieving a rate of 70%:
sup1.similarity(sup2)
0.7079433112862716
As you can see, their similarity score is quite high. This simple averaging method
works pretty decently. It is not, however, able to capture things such as negations
or a single deviating vector, which might not influence the average too much.
For example, "I would like to close a checking account" has a semantically different
meaning than, "I would like to open a checking account." However, the model
sees them as being pretty similar. Yet, this approach is still useful and a good
illustration of the advantages of representing semantics as vectors.
GoogLeNet Architecture from Szegedy and others' "Going Deeper with Convolutions"
Take a look at the preceding GoogLeNet architecture. While the graph is very
detailed, what we need to take away is the fact that the model is not just a number of
layers stacked on top of each other. Instead, there are multiple layers in parallel; in
this case, the model has three outputs. However, the question remains, how did the
authors build this complicated model? The sequential API wouldn't have allowed
them to, but the functional API makes it easy to string up layers like a pearl string
and create architectures such as the preceding one.
For many NLP applications, we need more complex models in which, for example,
two separate layers run in parallel. In the Keras functional API, we have more
control and can specify how layers should be connected. We can use this to create
much more advanced and complex models.
[ 205 ]
Parsing Textual Data with Natural Language Processing
We will use the functional API a lot more from now on. This section of the chapter
aims to provide a brief overview of the Keras functional API, as we will be going into
much more depth in later chapters. Firstly, let's look at a simple two-layer network in
both the sequential and functional way:
from keras.models import Sequential
from keras.layers import Dense, Activation
model = Sequential()
model.add(Dense(64, input_dim=64))
model.add(Activation('relu'))
model.add(Dense(4))
model.add(Activation('softmax'))
model.summary()
The preceding model is a simple model implemented in the sequential API. Take
note that this is how we have done it throughout this book so far. We will now
implement the same model in the functional API:
from keras.models import Model #1
from keras.layers import Dense, Activation, Input
model_input = Input(shape=(64,)) #2
x = Dense(64)(model_input) #3
x = Activation('relu')(x) #4
x = Dense(4)(x)
model_output = Activation('softmax')(x)
[ 206 ]
Chapter 5
1. Instead of defining the model first with model = Sequential(), you now
define the computational graph first and then turn it into a model using the
Model class.
2. Inputs are now their own layer.
3. Instead of using model.add(), you define the layer and then pass on an
input layer or the output tensor of the previous layer.
4. You create models by stringing layers on a chain. Dense(64)(model_input),
for instance, returns a tensor. You pass on this tensor to the next layer,
like in Activation('relu')(x). This function will return a new output
tensor, which you can pass to the next layer, and so on. This way, you create
a computational graph like a chain.
5. To create a model, you pass the model input layer as well as the final output
tensor of your graph into the Model class.
Functional API models can be used just like sequential API models. In fact, from the
output of this model's summary, you can see it is pretty much the same as the model
we just created with the sequential API:
Layer (type) Output Shape Param #
=================================================================
input_2 (InputLayer) (None, 64) 0
_________________________________________________________________
dense_3 (Dense) (None, 64) 4160
_________________________________________________________________
activation_3 (Activation) (None, 64) 0
_________________________________________________________________
dense_4 (Dense) (None, 4) 260
_________________________________________________________________
activation_4 (Activation) (None, 4) 0
=================================================================
Total params: 4,420
Trainable params: 4,420
Non-trainable params: 0
_________________________________________________________________
You can see that the functional API can connect layers in more advanced ways
than the sequential API. We can also separate the layer creation and connection step.
This keeps the code clean and allows us to use the same layer for different purposes.
[ 207 ]
Parsing Textual Data with Natural Language Processing
The following code segment will create the exact same model as the preceding
segment, but with separate layer creation and connection steps:
model_input = Input(shape=(64,))
dense = Dense(64)
x = dense(model_input)
activation = Activation('relu')
x = activation(x)
dense_2 = Dense(4)
x = dense_2(x)
model_output = Activation('softmax')(x)
Layers can be reused. For example, we could train some layers in one computational
graph and then use them for another, as we will do in the section on seq2seq models
later in the chapter.
One more caveat before we move on to use the functional API to build advanced
models. We should note that the activation function of any layer can also be specified
directly in the layer. So far, we have used a separate activation layer, which increases
clarity but is not strictly required. A Dense layer with a relu activation function can
also be specified as:
Dense(24, activation='relu')
When using the functional API, this can be easier than adding an activation function.
[ 208 ]
Chapter 5
Attention
Are you paying attention? If so, certainly not to everyone equally. In any text,
some words matter more than others. An attention mechanism is a way for a neural
network to focus on a certain element in a sequence. Focusing, for neural networks,
means amplifying what is important:
Attention layers are fully connected layers that take in a sequence and output
the weighting for a sequence. The sequence is then multiplied with the weightings:
def attention_3d_block(inputs,time_steps,
single_attention_vector = False):
input_dim = int(inputs.shape[2]) #1
a = Permute((2, 1),name='Attent_Permute')(inputs) #2
a = Reshape((input_dim, time_steps),name='Reshape')(a) #3
a = Dense(time_steps, activation='softmax',
name='Attent_Dense')(a) # Create attention vector #4
if single_attention_vector: #5
a = Lambda(lambda x: K.mean(x, axis=1),
name='Dim_reduction')(a) #6
a = RepeatVector(input_dim, name='Repeat')(a) #7
a_probs = Permute((2, 1), name='Attention_vec')(a) #8
output_attention_mul = Multiply(name='Attention_mul')
([inputs, a_probs]) #9
return output_attention_mul
[ 209 ]
Parsing Textual Data with Natural Language Processing
Let's break down the sequence we've just created. As you can see, it's made up of
nine key elements:
Attention block
[ 210 ]
Chapter 5
Notice how the preceding function defines takes a tensor as an input, defines a
graph, and returns a tensor. We can now call this function as part of our model
building process:
input_tokens = Input(shape=(maxlen,),name='input')
embedding = Embedding(max_words,
embedding_dim,
input_length=maxlen,
weights = [embedding_matrix],
trainable = False,
name='embedding')(input_tokens)
attention_mul = Flatten(name='flatten')(attention_mul)
output = Dense(1, activation='sigmoid',name='output')
(attention_mul)
model = Model(input_tokens, output)
In this case, we are using the attention block right after the embeddings. This means
that we can amplify or suppress certain word embeddings. Equally, we could use the
attention block after the LSTM. In many cases, you will find attention blocks to be
powerful tools in your arsenal when it comes to building models that deal with any
kind of sequence, especially in NLP.
To become more comfortable with how the functional API strings up layers and
how the attention block reshapes tensors, take a look at this model summary:
model.summary()
_________________________________________________________________________
_________________________
Layer (type) Output Shape Param #
Connected to
=========================================================================
=========================
input (InputLayer) (None, 140) 0
_________________________________________________________________________
_________________________
embedding (Embedding) (None, 140, 50) 500000 input[0]
[ 211 ]
Parsing Textual Data with Natural Language Processing
[0]
_________________________________________________________________________
_________________________
Attent_Permute (Permute) (None, 50, 140) 0
embedding[0][0]
_________________________________________________________________________
_________________________
Reshape (Reshape) (None, 50, 140) 0 Attent_
Permute[0][0]
_________________________________________________________________________
_________________________
Attent_Dense (Dense) (None, 50, 140) 19740
Reshape[0][0]
_________________________________________________________________________
_________________________
Dim_reduction (Lambda) (None, 140) 0 Attent_
Dense[0][0]
_________________________________________________________________________
_________________________
Repeat (RepeatVector) (None, 50, 140) 0 Dim_
reduction[0][0]
_________________________________________________________________________
_________________________
Attention_vec (Permute) (None, 140, 50) 0
Repeat[0][0]
_________________________________________________________________________
_________________________
Attention_mul (Multiply) (None, 140, 50) 0
embedding[0][0]
Attention_vec[0][0]
_________________________________________________________________________
_________________________
flatten (Flatten) (None, 7000) 0
Attention_mul[0][0]
_________________________________________________________________________
_________________________
output (Dense) (None, 1) 7001
flatten[0][0]
=========================================================================
=========================
Total params: 526,741
Trainable params: 26,741
Non-trainable params: 500,000
_________________________________________________________________________
_________________________
This model can be trained, just as any Keras model can be, and achieves around
80% accuracy on the validation set.
[ 212 ]
Chapter 5
Seq2seq models
In 2016, Google announced that it had replaced the entire Google Translate algorithm
with a single neural network. The special thing about the Google Neural Machine
Translation system is that it translates mutliple languages "end-to-end" using only
a single model. It works by encoding the semantics of a sentence and then decoding
the semantics into the desired output language.
The fact that such a system is possible at all baffled many linguists and other
researchers, as it shows that machine learning can create systems that accurately
capture high-level meanings and semantics without being given any explicit rules.
These semantic meanings are represented as an encoding vector, and while we don't
quite yet know how to interpret these vectors, there are a lot of useful applications for
them. Translating from one language to another is one such popular method, but we
could use a similar approach to "translate" a report into a summary. Text summarization
has made great strides, but the downside is that it requires a lot of computing power
to deliver meaningful results, so we will be focusing on language translation.
To deal with different lengths of phrases, we'll need to create an encoder that
aims to capture the sentence's semantic meaning. We then create a decoder that
has two inputs: the encoded semantics and the sequence that was already produced.
The decoder then predicts the next item in the sequence. For our character-level
translator, it looks like this:
[ 213 ]
Parsing Textual Data with Natural Language Processing
Note how the output of the decoder is used as the input of the decoder again. This
process is only stopped once the decoder produces a <STOP> tag, which indicates
that the sequence is over.
Note: The data and code for this section can be found on Kaggle at
https://www.kaggle.com/jannesklaas/a-simple-seq2seq-
translator.
The data
We use a dataset of English phrases and their translation. This dataset was obtained
from the Tabotea project, a translation database, and you can find the file attached
to the code on Kaggle. We implement this model on a character level, which means
that unlike previous models, we won't tokenize words, but characters. This makes
the task harder for our network because it now has to also learn how to spell words!
However, on the other hand, there are a lot fewer characters than words, therefore
we can just one-hot encode characters instead of having to work with embeddings.
This makes our model a bit simpler.
Input (English) and target (French) is tab delimited in the data file. Each row
represents a new phrase. The translations are separated by a tab (escaped character:
\t). So, we loop over the lines and read out inputs and targets by splitting the lines
at the tab symbol.
[ 214 ]
Chapter 5
To build up our tokenizer, we also need to know which characters are present in our
dataset. So, for all of the characters, we need to check whether they are already in our
set of seen characters, and if not, add them to it.
To do this, we must first set up the holding variables for texts and characters:
input_texts = []
target_texts = []
input_characters = set()
target_characters = set()
Then we loop over as many lines as we want samples and extract the texts and
characters:
lines = open(data_path).read().split('\n')
for line in lines[: min(num_samples, len(lines) - 1)]:
Let's break this code down so that we can understand it in more detail:
1. Input and target are split by tabs, English TAB French, so we split the
lines by tabs to obtain input and target texts.
2. We use \t as the "start sequence" character for the targets, and \n as
"end sequence" character. This way, we know when to stop decoding.
3. We loop over the characters in the input text, adding all characters that
we have not seen yet to our set of input characters.
4. We loop over the characters in the output text, adding all characters that
we have not seen yet to our set of output characters.
[ 215 ]
Parsing Textual Data with Natural Language Processing
Encoding characters
We now need to create lists of alphabetically sorted input and output characters,
which we can do by running:
input_characters = sorted(list(input_characters))
target_characters = sorted(list(target_characters))
We're also going to count how many input and output characters we have. This is
important since we need to know how many dimensions our one-hot encodings
should have. We can find this by writing the following:
num_encoder_tokens = len(input_characters)
num_decoder_tokens = len(target_characters)
Instead of using the Keras tokenizer, we will build our own dictionary mapping
characters to token numbers. We can do this by running the following:
input_token_index = {char: i for i,
char in enumerate(input_characters)}
target_token_index = {char: i for i,
char in enumerate(target_characters)}
We can see how this works by printing the token numbers for all characters in
a short sentence:
for c in 'the cat sits on the mat':
print(input_token_index[c], end = ' ')
63 51 48 0 46 44 63 0 62 52 63 62 0 58 57 0 63 51 48 0 56 44 63
Next, we build up our model training data. Remember that our model has two inputs
but only one output. While our model can handle sequences of any length, it is handy
to prepare the data in NumPy and thus to know how long our longest sequence is:
max_encoder_seq_length = max([len(txt) for txt in input_texts])
max_decoder_seq_length = max([len(txt) for txt in target_texts])
Now we prepare input and output data for our model. encoder_input_data is a
3D array of shape (num_pairs, max_english_sentence_length, num_english_
characters) containing a one-hot vectorization of the English sentences:
encoder_input_data = np.zeros((len(input_texts),
max_encoder_seq_length, num_encoder_tokens),dtype='float32')
[ 216 ]
Chapter 5
decoder_target_data = np.zeros((len(input_texts),
max_decoder_seq_length, num_decoder_tokens),dtype='float32')
You can see that the input and output of the decoder are the same except that the
output is one timestep ahead. This makes sense when you consider that we feed an
unfinished sequence into the decoder and want it to predict the next character. We
will use the functional API to create a model with two inputs.
You can see that the decoder also has two inputs: the decoder inputs and the encoded
semantics. The encoded semantics, however, are not directly the outputs of the
encoder LSTM but its states. In an LSTM, states are the hidden memory of the cells.
What happens is that the first "memory" of our decoder is the encoded semantics.
To give the decoder this first memory, we can initialize its states with the states of
the decoder LSTM.
[ 217 ]
Parsing Textual Data with Natural Language Processing
Now we define the decoder. The decoder uses the states of the encoder as initial
states for its decoding LSTM.
You can think of it like this: imagine you were a translator translating English to
French. When tasked with translating, you would first listen to the English speaker
and form ideas about what the speaker wants to say in your head. You would then
use these ideas to form a French sentence expressing the same idea.
It is important to understand that we are not just passing a variable, but a piece
of the computational graph. This means that we can later backpropagate from the
decoder to the encoder. In the case of our previous analogy, you might think that
your French translation suffered from a poor understanding of the English sentence,
so you might start changing your English comprehension based on the outcomes of
your French translation, for example:
decoder_inputs = Input(shape=(None, num_decoder_tokens),
name = 'decoder_inputs') #1
decoder_lstm = CuDNNLSTM(latent_dim,
return_sequences=True,
return_state=True,
name = 'decoder_lstm') #2
decoder_outputs, _, _ = decoder_lstm(decoder_inputs,
initial_state=encoder_states) #3
decoder_dense = Dense(num_decoder_tokens,
activation='softmax',
name = 'decoder_dense')
decoder_outputs = decoder_dense(decoder_outputs) #4
[ 218 ]
Chapter 5
We now have the pieces we need to define our model with two inputs and one output:
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
If you have the graphviz library installed, you can visualize the model very nicely
using the following code lines. Unfortunately, however, this code snippet won't
work on Kaggle:
from IPython.display import SVG
from keras.utils.vis_utils import model_to_dot
SVG(model_to_dot(model).create(prog='dot', format='svg'))
Seq2seq visual
[ 219 ]
Parsing Textual Data with Natural Language Processing
You can now compile and train the model. Since we have to choose between
a number of possible characters to output next, this is basically a multi-class
classification task. Therefore, we'll use a categorical cross-entropy loss:
model.compile(optimizer='rmsprop',
loss='categorical_crossentropy')
history = model.fit([encoder_input_data, decoder_input_data],
decoder_target_data,
batch_size=batch_size,
epochs=epochs,
validation_split=0.2)
The training process takes about 7 minutes on a GPU. However, if we were to plot
the model's progress, you can see that it's overfitting:
Seq2seq overfitting
The reason it's overfitting is largely because we used only 10,000 sentence pairs
of only relatively short sentences. To get a bigger model, a real translation or
summarization system would have to be trained on many more examples. To allow
you to follow the examples without owning a massive datacenter, we are just using
a smaller model to give an example of what a seq2seq architecture can do.
[ 220 ]
Chapter 5
The encoder model maps from the encoder inputs to the encoder states:
encoder_model = Model(encoder_inputs, encoder_states)
The decoder model then takes in the encoder memory plus its own memory from the
last character as an input. It then spits out a prediction plus its own memory to be
used for the next character:
#Inputs from the encoder
decoder_state_input_h = Input(shape=(latent_dim,)) #1
decoder_state_input_c = Input(shape=(latent_dim,))
#Decoder
decoder_outputs, state_h, state_c = decoder_lstm(decoder_inputs,
initial_state=decoder_states_inputs) #3
decoder_model = Model(
[decoder_inputs] + decoder_states_inputs,
[decoder_outputs] + decoder_states) #6
[ 221 ]
Parsing Textual Data with Natural Language Processing
1. The encoder memory consists of two states. We need to create two inputs
for both of them.
2. We then combine the two states into one memory representation.
3. We then connect the decoder LSTM we trained earlier to the decoder inputs
and the encoder memory.
4. We combine the two states of the decoder LSTM into one memory
representation.
5. We reuse the dense layer of the decoder to predict the next character.
6. Finally, we set up the decoder model to take in the character input as well as
the state input and map it to the character output as well as the state output.
Making translations
We can now start to use our model. To do this, we must first create an index that
maps tokens to characters again:
reverse_input_char_index = {i: char for char,
i in input_token_index.items()}
reverse_target_char_index = {i: char for char,
i in target_token_index.items()}
When we translate a phrase, we must first encode the input. We'll then loop, feeding
the decoder states back into the decoder until we receive a STOP; in our case, we use
the tab character to signal STOP.
states_value = encoder_model.predict(input_seq) #1
target_seq[0, 0, target_token_index['\t']] = 1. #3
stop_condition = False #4
decoded_sentence = ''
[ 222 ]
Chapter 5
output_tokens, h, c = decoder_model.predict(
[target_seq] + states_value) #6
sampled_char = reverse_target_char_index[
sampled_token_index] #8
decoded_sentence += sampled_char #9
return decoded_sentence
For the final time in this chapter, let's break down the code:
[ 223 ]
Parsing Textual Data with Natural Language Processing
Now we can translate English into French! At least for some phrases, it works quite
well. Given that we did not supply our model with any rules about French words
or grammar, this is quite impressive. Translation systems such as Google Translate,
of course, use much bigger datasets and models, but the underlying principles are
the same.
We then one-hot encode all characters in the text by setting the element at the index
of the characters' token numbers to 1:
for i, char in enumerate(my_text):
print(i,char, input_token_index[char])
placeholder[0,i,input_token_index[char]] = 1
This will print out the characters' token numbers alongside the character and its
position in the text:
0 T 38
1 h 51
2 a 44
3 n 57
4 k 54
5 s 62
6 ! 1
Seq2seq models are useful not only for translating between languages. They can be
trained on just about anything that takes a sequence as an input and also outputs
a sequence.
Remember our forecasting task from the last chapter? The winning solution to the
forecasting problem was a seq2seq model. Text summarization is another useful
application. Seq2seq models can also be trained to output a series of actions,
such as a sequence of trades that would minimize the impact of a large order.
[ 224 ]
Chapter 5
Exercises
Now that we're at the end of the chapter, let's see what we've learned. To finish this
chapter, I've included three exercises that will challenge you based on what we've
covered in this chapter:
1. Add an extra layer to the encoder of the translation model. The translation
model might work better if it had a bit more capacity to learn the structure
of French sentences. Adding one more LSTM layer will be a good exercise
to learn about the functional API.
2. Add attention to the encoder of the translation model. Attention will allow
the model to focus on the (English) words that really matter for translation.
It is best to use attention as the last layer. This task is a bit harder than the
previous one, but you will understand the inner workings of attention much
better.
3. Visit Daily News for Stock Market Prediction at https://www.kaggle.com/
aaron7sun/stocknews. The task is to use the daily news as an input to
predict stock prices. There are a number of kernels already that can help
you with this. Use what you have learned in this chapter to predict some
stock prices!
Summary
In this chapter, you have learned the most important NLP techniques. There was
a lot that we've learned, and here's a big list of things we covered in this chapter
and everything you should now feel confident about understanding:
[ 225 ]
Parsing Textual Data with Natural Language Processing
You now have a big set of tools in your toolbox that will allow you to tackle NLP
problems. Throughout the rest of this book, you will see some of these techniques
again, being used in different contexts to solve hard problems. These techniques
are useful across the industry, from retail banking to hedge fund investing. While
the problem your institution is trying to solve might require a bit of tweaking, the
general approaches are quite transferable.
In the next chapter, we will look at a technique that has gained a lot of attention since
DeepMind beat a human Go champion: reinforcement learning. This technique is
especially useful when working in financial markets and is in many ways a natural
extension of what many quantitative investment firms are already doing. So, stay
tuned, and I'll see you on the other side.
[ 226 ]
Using Generative Models
Generative models generate new data. In a way, they are the exact opposite of the
models that we've dealt with in prior chapters. While an image classifier takes in
a high-dimensional input, the image, and outputs a low-dimensional output such
as the content of the image, a generative model goes about things in exactly the
opposite way around. It might, for example, draw images from the description
of what's in them.
Generative models are still in the experimental phase of their development, and are
currently used mostly in image applications. However, they are an important model
as shown by the fact that there have already been several applications that have used
generative models that have caused an uproar within the industry.
This technology is not completely negative, there are positive applications as well,
especially if the generative model's data is sparse. If this is the case, generative
models can generate realistic data that other models can then train on. Generative
models are able to "translate" images, a prime example being taking satellite
images and turning them into street maps. Another example is that generative
models can generate code from website screenshots. They can even be used to
combat unfairness and discrimination in machine learning models, as we will
see in Chapter 9, Fighting Bias.
[ 227 ]
Using Generative Models
In the field of finance, data is frequently sparse. Think back to the fraud case from
Chapter 2, Applying Machine Learning to Structured Data, in which we were classifying
fraudulent transactions from transaction metadata. We found that there was not
much fraud taking place in the dataset that we used, so the model had a hard time
detecting when fraud was taking place. Usually, when this occurs, engineers make
assumptions and create synthetic data. Machine learning models, however, can do
this themselves, and in the process, they might even discover some useful features
that can help with fraud detection.
Understanding autoencoders
Technically, autoencoders are not generative models since they cannot create
completely new kinds of data. Yet, variational autoencoders, a minor tweak to
vanilla autoencoders, can. So, it makes sense to first understand autoencoders
by themselves, before adding the generative element.
x = f ( x)
This might sound trivial at first, but the trick here is that autoencoders have
a bottleneck. The middle hidden layer's size is smaller than the size of the input,
x. Therefore, the model has to learn a compressed representation that captures all
of the important elements of x in a smaller vector.
[ 228 ]
Chapter 6
This can best be shown in the following diagram, where we can see a compressed
representation of the Autoencoder scheme:
Autoencoder scheme
This compressed representation aims to capture the essence of the input, which turns
out to be useful for us. We might, for example, want to capture what essentially
distinguishes a fraudulent transaction from a genuine one. Vanilla autoencoders
accomplish this with something similar to standard principal component analysis
(PCA). They allow us to reduce the dimensionality of our data and focus on what
matters. But in contrast to PCA, autoencoders can be extended in order to generate
more data of a certain type. For example, autoencoders can better deal with image
or video data since they can make use of the spatiality of data using convolutional
layers.
In this section, we will build two autoencoders. The first will be used for handwritten
digits from the MNIST dataset. Generative models are easier to debug and
understand for visual data due to the fact that humans are intuitively good at
judging whether two pictures show something similar, but are less good at judging
abstract data. The second autoencoder is for a fraud detection task, using similar
methods as the MNIST dataset.
Before diving into the code described here, make sure you have saved the MNIST
dataset on the right path, successfully imported both the NumPy and Matplotlib
libraries, and set a random seed to ensure that your experiments are reproducible.
[ 229 ]
Using Generative Models
Note: You can find the code for the MNIST autoencoder and variational
autoencoder under the following URL https://www.kaggle.com/
jannesklaas/mnist-autoencoder-vae.
We're going to set the encoding dimensionality hyperparameter now so that we can
use it later:
encoding_dim = 32
Then, we construct the autoencoder using the Keras functional API. While a simple
autoencoder could be constructed using the sequential API, this is a good refresher
for us on how the functional API works.
First, we import the Model class, which allows us to create functional API models.
We also need to import both the Input and Dense layers. You'll remember from
previous chapters how the functional API needs a separate input layer, while
the sequential API does not need one. To import both layers, we need to run
the following:
from keras.models import Model
from keras.layers import Input, Dense
This is followed by a Dense decoding layer that aims to reconstruct the original
image:
input_img = Input(shape=(784,))
After we have created and chained up the layers, we are then able to create a model
that maps from the input to the decoded image:
autoencoder = Model(input_img, decoded)
To get a better idea of what is going on, we can plot a visualization of the resulting
autoencoder model with the following code:
from keras.utils import plot_model
plot_model(autoencoder, to_file='model.png', show_shapes=True) plt.
figure(figsize=(10,10))
plt.imshow(plt.imread('model.png'))
[ 230 ]
Chapter 6
Autoencoder model
To train this autoencoder, we use the X values as both the input and output:
autoencoder.fit(X_train_flat, X_train_flat,
epochs=50,
batch_size=256,
shuffle=True,
validation_data=(X_test_flat, X_test_flat))
After we train this autoencoder, which will take between one and two minutes,
we can visually inspect how well it is doing. To do this, we first extract a single
image from the test set, before adding a batch dimension to the image in order to
run it through the model, which is what we use np.expand_dims for:
original = np.expand_dims(X_test_flat[0],0)
Now we're going to run the original image through the autoencoder. You'll
remember that the original MNIST image showed us a number seven, so we're
hoping that the output of our autoencoder shows a seven as well:
seven = autoencoder.predict(original)
Next, we're going to reshape both the autoencoder output as well as the original
image back into 28x28-pixel images:
seven = seven.reshape(1,28,28)
original = original.reshape(1,28,28)
We then plot the original and reconstructed image next to each other. matplotlib
does not allow the image to have a batch dimension, therefore we need to pass an
array without it. By indexing the images with [0,:,:], we'll only pass the first item
in the batch with all pixels.
[ 231 ]
Using Generative Models
b=fig.add_subplot(1,2,2)
b.set_title('Autoencoder')
imgplot = plt.imshow(seven[0,:,:])
After running that code, you'll see that our hopes have been achieved! Compared
to the original image (left), our autoencoder image (right) is also showing a seven!:
Autoencoder result
As you can see in the preceding screenshot, the reconstructed seven is still a seven,
so the autoencoder was able to capture the general idea of what a seven is. It's not
perfect though, as you can see it's a bit blurry around the edges, especially in the
top left. It seems that while the autoencoder is unsure about the length of the lines,
it does have a good idea that there are two lines in a seven and it is aware of the
general direction they follow.
An autoencoder such as this one performs nonlinear PCA. It learns which
components matter the most for a seven to be a seven. The usefulness of being able
to learn this representation goes beyond images. Within credit card fraud detection,
such principal components would make for good features that another classifier
would be able to work with.
In the next section, we will apply an autoencoder to the credit card fraud problem.
This new dataset contains records of actual credit card transactions with anonymized
features; however, it does not lend itself much to feature engineering. Therefore, we will
have to rely on end-to-end learning methods in order to build a good fraud detector.
As usual, we first load the data. The Time feature shows the absolute time of the
transaction, which makes the data a bit hard to deal with here. Therefore, we will
just drop it, which we can do by running:
df = pd.read_csv('../input/creditcard.csv')
df = df.drop('Time',axis=1)
We then separate the X data on the transaction from the classification of the
transaction and extract the NumPy array that underlies the pandas DataFrame:
X = df.drop('Class',axis=1).values
y = df['Class'].values
Now we need to scale the features. Feature scaling makes it easier for our model
to learn a good representation of the data. This time around, we're going employ
a slightly different method of feature scaling than what we did before. We'll scale
all features to be between zero and one, as opposed to having a mean of zero and
a standard deviation of one. By doing this, we ensure that there are neither any
very high nor very low values in the dataset.
We must be aware that this method is susceptible to outliers influencing the result.
For each column, we first subtract the minimum value, so that the new minimum
value becomes zero. Next, we divide by the maximum value so that the new
maximum value becomes one.
[ 233 ]
Using Generative Models
We then create the exact same autoencoder as we did before; however, this
time, we do it with different dimensions. Our input now has 29 dimensions,
which we compress down to 12 dimensions before aiming to restore the original
29-dimensional output.
We are going to use the sigmoid activation function for the decoded data. This
is only possible because we've scaled the data to have values between zero and one.
We are also using a tanh activation within the encoded layer. This is just a style
choice that worked well in experiments and ensures that encoded values are all
between minus one and one. With that being said, you may use different activation
functions depending on your individual needs.
If you are working with images or deeper networks, a ReLU activation is usually
a good choice. However, if you are working with a shallower network, as we are
doing here, then a tanh activation often works well:
data_in = Input(shape=(29,))
encoded = Dense(12,activation='tanh')(data_in)
decoded = Dense(29,activation='sigmoid')(encoded)
autoencoder = Model(data_in,decoded)
In this example, we've used a mean squared error loss. This seems a bit of an
unusual choice at first, using a sigmoid activation with a mean squared error loss,
yet it makes sense. Most people think that sigmoid activations have to be used with
a cross-entropy loss, but cross-entropy loss encourages values to either be zero or
one, which works well for classification tasks where this is the case.
In our credit card example, most values will be around 0.5. Mean squared error,
which we can see being implemented in the code below, is better at dealing with
values where the target is not binary, but on a spectrum. Binary cross entropy forces
values to be close to zero and one, which is not what we always want:
autoencoder.compile(optimizer='adam',loss='mean_squared_error')
After training, which will take around two minutes, the autoencoder converges
to a low loss:
autoencoder.fit(X_train,
X_train,
epochs = 20,
batch_size=128,
validation_data=(X_test,X_test))
[ 234 ]
Chapter 6
The reconstruction loss is low, but how do we know whether our autoencoder is
working well? Once again, a visual inspection will come to the rescue. As we've
explained before, humans are very good at judging things visually, but not very
good at judging abstract numbers.
To run a visual inspection, first we must make some predictions, in which we'll run
a subset of our test set through the autoencoder:
pred = autoencoder.predict(X_test[0:10])
We must can then plot individual samples. The following code produces an
overlaid bar chart comparing the original transaction data with the reconstructed
transaction data:
import matplotlib.pyplot as plt
import numpy as np
width = 0.8
prediction = pred[9]
true_value = X_test[9]
indices = np.arange(len(prediction))
fig = plt.figure(figsize=(10,7))
plt.xticks(indices+width/2.,
['V{}'.format(i) for i in range(len(prediction))] )
plt.legend()
plt.show()
[ 235 ]
Using Generative Models
As you can see, our model does a fine job at reconstructing the original values.
The reconstructed values often match the true values, and if they don't, then they
only deviate by a small margin. As you can see, visual inspection gives more insight
than looking at abstract numbers.
To create the encoder model, all we have to do is to define a new Keras model
that maps from the input to the encoded state:
encoder = Model(data_in,encoded)
Note that you don't need to train this model again. The layers keep the weights
from the previously trained autoencoder.
[ 236 ]
Chapter 6
But how would we know whether these encodings contain any meaningful
information about fraud? Once again, visual representation is key. While our
encodings have fewer dimensions than the input data, they still have 12 dimensions.
It's impossible for humans to think about a 12-dimensional space, so we need
to draw our encodings in a lower dimensional space while still preserving the
characteristics we care about.
In our case, the characteristic we care about is proximity. We want points that are
close to each other in the 12-dimensional space to be close to each other in the
2-dimensional plot. More precisely, we care about the neighborhood. We want the
points that are closest to each other in the high-dimensional space to also be closest
to each other in the low-dimensional space.
In a nutshell, t-SNE aims to faithfully represent the probability that two points are
neighbors in a random sample of all points. That is, it tries to find a low-dimensional
representation of data in which points in a random sample have the same probability
of being the closest neighbors as in the high-dimensional data:
[ 237 ]
Using Generative Models
pi| j =
(
exp − xi − x j
2
2σi 2 )
∑ k ≠i (
exp − xi − xk
2
2σi 2 )
pi| j + p j ,i
pij =
2n
qij =
(1+ yi − y j )
2 −1
∑ (1 + )
2 −1
k ≠l
yk − yl
4. Just like in training neural networks, we will optimize the positions of the
data points in the lower dimensional space by following the gradient of
a loss function. The loss function, in this case, is the Kullback–Leibler (KL)
divergence between the similarities in the higher and lower dimensional
space. We will give the KL divergence a closer look in the section on
variational autoencoders. For now, just think of it as a way to measure the
difference between two distributions. The derivative of the loss function with
respect to the position yi of data point i in the lower dimensional space is
as follows:
[ 238 ]
Chapter 6
dL
dyi
(
= 4∑ ( pij − qij )( yi − y j ) 1 + yi − y j
2 −1
)
5. Adjust the data points in the lower dimensional space by using gradient
descent, moving points that were close in the high-dimensional data closer
together and moving points that were further away further from each other:
dL
y( ) = y(
t t −1)
+
dy
(
+ α (t ) y( ) − y( )
t −1 t −2
)
6. You will recognize this as a form of gradient descent with momentum,
as the previous gradient is incorporated into the updated position.
The t-distribution used always has one degree of freedom. This freedom leads
to a simpler formula as well as some nice numerical properties that lead to faster
computation and more useful charts.
H ( Pi )
Perp ( Pi ) = 2
Here Pi is a probability distribution over the position of all data points in the dataset
and H ( Pi ) is the Shanon entropy of this distribution, calculated as follows:
While the details of this formula are not very relevant to using t-SNE, it is important
to know that t-SNE performs a search over values of the standard deviation, σ , so that
it finds a global distribution, Pi , for which the entropy over our data is of our desired
perplexity. In other words, you need to specify the perplexity by hand, but what that
perplexity means for your dataset also depends on the dataset itself.
Laurens Van Maarten and Geoffrey Hinton, the inventors of t-SNE, report that the
algorithm is relatively robust for choices of perplexity between 5 and 50. The default
value in most libraries is 30, which is a fine value for most datasets. However, if you
find that your visualizations are not satisfactory, then tuning the perplexity value is
probably the first thing you would want to do.
[ 239 ]
Using Generative Models
For all the math involved, using t-SNE is surprisingly simple. Scikit-learn has
a handy t-SNE implementation that we can use just like any algorithm in scikit-learn.
We first import the TSNE class, and then we can create a new TSNE instance. We
define that we want to train for 5000 epochs, and use the default perplexity of 30 and
the default learning rate of 200. We also specify that we would like output during the
training process. We then call fit_transform, which transforms our 12 encodings
into 2-dimensional projections:
from sklearn.manifold import TSNE
tsne = TSNE(verbose=1,n_iter=5000)
res = tsne.fit_transform(enc)
There's also a faster Python implementation of t-SNE that can be used as a drop-
in replacement of the scikit-learn implementation. However, this is not as well
documented and contains fewer features, therefore we will not be covering it
in this book.
We can then plot our t-SNE results as a scatterplot. For illustration, we will
distinguish frauds from non-frauds by color, with frauds being plotted in red and
non-frauds being plotted in blue. Since the actual values of t-SNE do not matter as
much, we will hide the axes:
fig = plt.figure(figsize=(10,7))
scatter =plt.scatter(res[:,0],res[:,1],c=y_test,
cmap='coolwarm', s=0.6)
scatter.axes.get_xaxis().set_visible(False)
scatter.axes.get_yaxis().set_visible(False)
[ 240 ]
Chapter 6
Let's now see, what the output chart will look like:
For easier spotting, and for those reading the print version, the cluster containing
the most frauds, those that are marked red, has been marked with a circle. You can see
that the frauds are nicely separate from the rest of the genuine transactions, those in
blue. Clearly, our autoencoder has found a way to distinguish frauds from the genuine
transaction without being given labels. This is a form of unsupervised learning.
Using our autoencoder, we could now use the encoded information as features
for a classifier. However, what's even better is that with only a slight modification
of the autoencoder, we can generate more data that has the underlying properties
of a fraud case while having different features. This is done with a variational
autoencoder, which will be the focus of the next section.
[ 241 ]
Using Generative Models
Variational autoencoders
Autoencoders are basically an approximation for PCA. However, they can be
extended to become generative models. Given an input, variational autoencoders
(VAEs) can create encoding distributions. This means that for a fraud case, the
encoder would produce a distribution of possible encodings that all represent the
most important characteristics of the transaction. The decoder would then turn all
of the encodings back into the original transaction.
This is useful since it allows us to generate data about transactions. One problem of
fraud detection that we discovered earlier is that there are not all that many fraudulent
transactions. Therefore, by using a VAE, we can sample any amount of transaction
encodings and train our classifier with more fraudulent transaction data.
So, how do VAEs do it? Instead of having just one compressed representation vector,
a VAE has two: one for the mean encoding, µ , and one for the standard deviation of
this encoding, σ :
VAE scheme
Both the mean and standard deviation are vectors, just as with the encoding vector
we used for the vanilla autoencoder. However, to create the actual encoding,
we simply need to add random noise with the standard deviation, σ , to our
encoding vector.
[ 242 ]
Chapter 6
MNIST example
Now on to our first VAE. This VAE will work with the MNIST dataset and give you
a better idea about how VAEs work. In the next section, we will build the same VAE
for credit card fraud detection.
Notice the two new imports, the Lambda layer and the metrics module. The metrics
module provides metrics, such as the cross-entropy loss, which we will use to build
our custom loss function. Meanwhile the Lambda layer allows us to use Python
functions as layers, which we will use to sample from the encoding distribution.
We will see just how the Lambda layer works in a bit, but first, we need to set up
the rest of the neural network.
The first thing we need to do is to define a few hyperparameters. Our data has
an original dimensionality of 784, which we compress into a latent vector with 32
dimensions. Our network has an intermediate layer between the input and the latent
vector, which has 256 dimensions. We will train for 50 epochs with a batch size of 100:
batch_size = 100
original_dim = 784
latent_dim = 32
intermediate_dim = 256
epochs = 50
For computational reasons, it is easier to learn the log of the standard deviation
rather than the standard deviation itself. To do this we create the first half of
our network, in which the input, x, maps to the intermediate layer, h. From this
layer, our network splits into z_mean, which expresses µ and z_log_var, which
expresses log σ :
x = Input(shape=(original_dim,))
h = Dense(intermediate_dim, activation='relu')(x)
z_mean = Dense(latent_dim)(h)
z_log_var = Dense(latent_dim)(h)
[ 243 ]
Using Generative Models
Additionally, a Lambda function can only take one input argument. In the layer we
want to create, the input is just the previous layer's output tensor. In this case, we
want to create a layer with two inputs, µ and σ . Therefore, we will wrap both inputs
into a tuple that we can then take apart.
1. We take apart the input tuple and have our two input tensors.
2. We create a tensor containing random, normally distributed noise with
a mean of zero and a standard deviation of one. The tensor has the shape
as our input tensors (batch_size, latent_dim).
3. Finally, we multiply the random noise with our standard deviation to give
it the learned standard deviation and add the learned mean. Since we are
learning the log standard deviation, we have to apply the exponent function
to our learned tensor.
All these operations are differentiable since we are using the Keras backend
functions. Now we can turn this function into a layer and connect it to the previous
two layers with one line:
z = Lambda(sampling)([z_mean, z_log_var])
[ 244 ]
Chapter 6
And voilà! We've now got a custom layer that samples from a normal distribution
described by two tensors. Keras can automatically backpropagate through this layer
and train the weights of the layers before it.
Now that we have encoded our data, we also need to decode it as well. We are able
to do this with two Dense layers:
decoder_h = Dense(intermediate_dim, activation='relu')(z)
x_decoded = Dense(original_dim, activation='sigmoid')
decoder_mean(h_decoded)
Our network is now complete. This network will encode any MNIST image
into a mean and a standard deviation tensor from which the decoding part then
reconstructs the image. The only thing missing is the custom loss incentivizing the
network to both reconstruct images and produce a normal Gaussian distribution in
its encodings. Let's address that now.
Kullback–Leibler divergence
To create the custom loss for our VAE, we need a custom loss function. This
loss function will be based on the Kullback-Leibler (KL) divergence.
KL divergence, is one of the metrics, just like cross-entropy, that machine learning
inherited from information theory. While it is used frequently, there are many
struggles you can encounter when trying to understand it.
At its core, KL divergence measures how much information is lost when distribution
p is approximated with distribution q.
Imagine you are working on a financial model and have collected data on the
returns of a security investment. Your financial modeling tools all assume a normal
distribution of returns. The following chart shows the actual distribution of returns
versus an approximation using a normal distribution model. For the sake of this
example, let's assume there are only discrete returns. Before we go ahead, be assured
that we'll cover continuous distributions later:
[ 245 ]
Using Generative Models
Of course, the returns in your data are not exactly normally distributed. So, just how
much information about returns would you lose if you did lose the approximation?
This is exactly what the KL divergence is measuring:
N
DKL ( p q) = ∑ p ( xi )⋅ (log p ( xi ) − log q ( xi ))
i =1
Here p ( xi ) and q ( xi ) are the probabilities that x, in this case, the return, has some
value i, say 5%. The preceding formula effectively expresses the expected difference
in the logarithm of probabilities of the distributions p and q:
This expected difference of log probabilities is the same as the average information
lost if you approximate distribution p with distribution q. See the following:
a
log a − log b = log
b
N
p ( xi )
DKL ( p q) = ∑ p ( xi )⋅ log
i =1 q ( xi )
[ 246 ]
Chapter 6
When p is substituted with the normal Gaussian distribution, N (0,1) , and the
approximation q is a normal distribution with a mean of µ and a standard
deviation of σ , N (µ, σ ) , the KL divergence, simplifies to the following:
The partial derivatives to our mean and standard deviation vectors are, therefore
as follows:
dDKL
=µ
dµ
You can see that the derivative with respect to µ is zero if µ is zero, and the
derivative with respect to σ is zero if σ is one. This loss term is added to the
reconstruction loss.
The reconstruction loss is the same loss that we applied for the vanilla autoencoder.
Binary cross-entropy is an appropriate loss for MNIST reconstruction. Since Keras'
implementation of a binary cross-entropy loss already takes the mean across the
batch, an operation we only want to do later, we have to scale the loss back up,
so that we can divide it by the output dimensionality:
reconstruction_loss = original_dim *
metrics.binary_crossentropy(x, x_decoded)
[ 247 ]
Using Generative Models
Expressed in Python, the KL divergence loss appears like the following code:
kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean)
- K.exp(z_log_var), axis=-1)
Our final loss is then the mean of the sum of the reconstruction loss and KL
divergence loss:
vae_loss = K.mean(reconstruction_loss + kl_loss)
Since we have used the Keras backend for all of the calculations, the resulting loss is a
tensor that can be automatically differentiated. Now we can create our model as usual:
vae = Model(x, x_decoded)
Since we are using a custom loss, we have the loss separately, and we can't just add it
in the compile statement:
vae.add_loss(vae_loss)
Now we will compile the model. Since our model already has a loss, we only have
to specify the optimizer:
vae.compile(optimizer='rmsprop')
Another side effect of the custom loss is that it compares the output of the VAE with the
input of the VAE, which makes sense as we want to reconstruct the input. Therefore,
we do not have to specify the y values, as only specifying an input is enough:
vae.fit(X_train_flat,
shuffle=True,
epochs=epochs,
batch_size=batch_size,
validation_data=(X_test_flat, None))
In the next section we will learn how we can use a VAE to generate data.
[ 248 ]
Chapter 6
To showcase this, from our test data, we're going to take a seven:
one_seven = X_test_flat[0]
We then add a batch dimension and repeat the seven across the batch four times.
After which we now have a batch of four, identical sevens:
one_seven = np.expand_dims(one_seven,0)
one_seven = one_seven.repeat(4,axis=0)
We can then make a prediction on that batch, in which case, we get back the
reconstructed sevens:
s = vae.predict(one_seven)
The next step is broken in two parts. Firstly, we're going to reshape all the sevens
back into image form:
s= s.reshape(4,28,28)
As a result of running the code that we've just walked through, we'll then see the
following screenshot showing our four sevens as our output:
A collection of sevens
[ 249 ]
Using Generative Models
As you can see, all of the images show a seven. While they look quite similar, if you
look closely, you can see that there are several distinct differences. The seven on the
top left has a less pronounced stroke than the seven on the bottom left. Meanwhile,
the seven on the bottom right has a sight bow at the end.
What we've just witnessed is the VAE successfully creating new data. While using
this data for more training is not as good as compared to using using completely new
real-world data, it is still very useful. While generative models such as this one are
nice on the eye we will now discuss how this technique can be used for credit card
fraud detection.
The following visualization shows the resulting VAE including both the input and
output shapes:
[ 250 ]
Chapter 6
Armed with a VAE that can encode and generate credit card data, we can now tackle
the task of an end-to-end fraud detection system. This can reduce bias in predictions
as we can learn complicated rules directly from data.
How would we detect when someone engages in spoofing? One way is to use an
autoencoder. By using a large amount of order book information, we can train an
autoencoder to reconstruct "normal" trading behavior. For traders whose trading
patterns deviate a lot from normal trading, the reconstruction loss of the trained
autoencoder for the transaction will be quite high.
Recurrent neural networks (RNNs), by default, take in a time series and output
a single vector. They can also output sequences if Keras' return_sequences
argument is set to True. Using recurrent neural networks such as LSTMs,
building an autoencoder for time series can be done using the following code:
from keras.models import Sequential
from keras.layers import LSTM, RepeatVector
model = Sequential() #1
model.add(LSTM(latent_dim, input_shape=(maxlen, nb_features))) #2
model.add(RepeatVector(maxlen)) #3
model.add(LSTM(nb_features, return_sequences=True)) #4
Let's pause for a second and break down what we've just coded. As you can see,
there are four key elements to this code:
[ 252 ]
Chapter 6
It is reasonable to assume that just because two market days look a bit different,
the same forces might be at work. Mathematically, we can assume that market data
{ x } is sampled from a probability distribution, p(x), with a small number of latent
k
This solves the problem that a standard maximum likelihood model for this kind of
problem is computationally intractable. Two other methods performing the same
feat are the Markov Chain Monte Carlo and Hamilton Monte Carlo methods. While
neither will be covered in depth here, though they will be featured in later chapters,
it's worth understanding that VAEs address long-standing problems in mathematical
finance in a computationally tractable way.
Generative models can also be used to solve problems beyond the scope of
traditional methods. Financial markets are fundamentally adversarial environments
in which investors are trying to achieve something that is impossible in aggregate:
above-average returns. Knowing that a company is doing well is not enough: if
everyone knows the company is doing well, then the stock price will be high and
returns will be low. The key is knowing that a company is doing well while everyone
else believes it is doing poorly. Markets are a zero-sum game-theoretic environment.
GANs make use of these dynamics to generate realistic data.
GANs
GANs work a lot like an art forger and a museum curator. Every day, the art
forger tries to sell some fake art to the museum, and every day the curator tries
to distinguish whether a certain piece is real or fake. The forger learns from their
failures. By trying to fool the curator and observing what leads to success and failure,
they become a better forger. But the curator learns too. By trying to stay ahead of the
forger, they become a better curator. As time passes, the forgeries become better and
so does the distinguishing process. After years of battle, the art forger is an expert
that can draw just as well as Picasso and the curator is an expert that can distinguish
a real painting by tiny details.
[ 253 ]
Using Generative Models
GAN scheme
Once again, generative models are easier to understand when images are generated,
so in this section, we will look at image data, although all kinds of data can be used.
[ 254 ]
Chapter 6
In this formula x is an item drawn from the distribution of real data, pdata, and z is
a latent vector drawn from the latent vector space, pz.
The output distribution of the generator is noted as pg. It can be shown that the global
optimum of this game is pg = pdata , that is, if the distribution of the generated data is
equal to the distribution of actual data.
GANs get optimized following a game-theoretic value function. Solving this type
of optimization problem with deep learning is an active area of research, and an
area we will visit again in Chapter 8, Privacy, Debugging, and Launching Your Products,
where we will discuss reinforcement learning. The fact that deep learning can be
used to solve Minimax games is exciting news for the field of finance and economics,
which features many such problems.
A MNIST GAN
Let's now implement a GAN in order to generate MNIST characters. Before we start,
we need to do some imports. GANs are large models, and in this section you will see
how to combine sequential and functional API models for easy model building:
from keras.models import Model, Sequential
[ 255 ]
Using Generative Models
• LeakyReLU is just like ReLU, except that the activation allows for small
negative values. This prevents the gradient from ever becoming zero. This
activation function works well for GANs, something we will discuss in the
next section:
Leaky ReLU
• Reshape does the same as np.reshape: it brings a tensor into a new form.
• UpSampling2D scales a 2D feature map up, for example, by a factor of two,
by repeating all numbers in the feature map.
Neural network layers get initialized randomly. Usually, the random numbers
are drawn from a distribution that supports learning well. For GANs, it turns
out that a normal Gaussian distribution is a better alternative:
from keras.initializers import RandomNormal
generator.add(Dense(128*7*7, input_dim=latent_dim,
kernel_initializer=RandomNormal(stddev=0.02))) #2
generator.add(LeakyReLU(0.2)) #3
generator.add(Reshape((128, 7, 7))) #4
generator.add(UpSampling2D(size=(2, 2))) #5
generator.add(Conv2D(64,kernel_size=(5, 5),padding='same')) #6
[ 256 ]
Chapter 6
generator.add(LeakyReLU(0.2)) #7
generator.add(UpSampling2D(size=(2, 2))) #8
Again, let's take a look at the generator model code, which consists of 10 key steps:
[ 257 ]
Using Generative Models
discriminator.add(LeakyReLU(0.2))
discriminator.add(Dropout(0.3))
discriminator.add(Conv2D(128, kernel_size=(5, 5),
strides=(2, 2),
padding='same'))
discriminator.add(LeakyReLU(0.2))
discriminator.add(Dropout(0.3)) #2
discriminator.add(Flatten())
discriminator.add(Dense(1, activation='sigmoid'))
discriminator.compile(loss='binary_crossentropy', optimizer=adam)
1. As with the generator, the first layer of the discriminator should be initialized
randomly from a Gaussian distribution.
2. Dropout is commonly used in image classifiers. For GANs, it should also
be used just before the last layer.
Now we have both a generator and a discriminator. To train the generator, we have
to get the gradients from the discriminator to backpropagate through and train the
generator. This is where the power of Keras' modular design comes into play.
The following code creates a GAN model that can be used to train the generator from
the discriminator gradients:
discriminator.trainable = False #1
ganInput = Input(shape=(latent_dim,)) #2
x = generator(ganInput) #3
ganOutput = discriminator(x) #4
gan = Model(inputs=ganInput, outputs=ganOutput) #5
gan.compile(loss='binary_crossentropy', optimizer=adam) #6
[ 258 ]
Chapter 6
Training our GAN requires some customization of the training process and a couple
of GAN-specific tricks as well. More specifically, we have to write our own training
loop, something that we'll achieve with the following code:
epochs=50
batchSize=128
batchCount = X_train.shape[0] // batchSize #1
noise = np.random.normal(0, 1,
size=[batchSize, latent_dim]) #4
imageBatch = X_train[np.random.randint(0,
X_train.shape[0],
size=batchSize)] #5
generatedImages = generator.predict(noise) #6
X = np.concatenate([imageBatch, generatedImages]) #7
yDis = np.zeros(2*batchSize) #8
yDis[:batchSize] = 0.9
[ 259 ]
Using Generative Models
labelNoise = np.random.random(yDis.shape) #9
yDis += 0.05 * labelNoise + 0.05
noise = np.random.normal(0, 1,
size=[batchSize, latent_dim]) #12
yGen = np.ones(batchSize) #13
discriminator.trainable = False #14
gloss = gan.train_on_batch(noise, yGen) #15
dLosses.append(dloss) #16
gLosses.append(gloss)
That was a lot of code we just introduced. So, let's take a minute to pause and think
about the 16 key steps:
1. We have to write a custom loop to loop over the batches. To know how many
batches there are, we need to make an integer division of our dataset size by
our batch size.
2. In the outer loop, we iterate over the number of epochs we want to train.
3. In the inner loop, we iterate over the number of batches we want to train on
in each epoch. The tqdm tool helps us keep track of progress within the batch.
4. We create a batch of random latent vectors.
5. We randomly sample a batch of real MNIST images.
6. We use the generator to generate a batch of fake MNIST images.
7. We stack the real and fake MNIST images together.
8. We create the target for our discriminator. Fake images are encoded with
0, and real images with 0.9. This technique is called soft labels. Instead of
hard labels (zero and one), we use something softer in order to not train the
GAN too aggressively. This technique has been shown to make GAN training
more stable.
[ 260 ]
Chapter 6
9. On top of using soft labels, we add some noise to the labels. This, once again,
will make the training more stable.
10. We make sure that the discriminator is trainable.
11. We train the discriminator on a batch of real and fake data.
12. We create some more random latent vectors for training the generator.
13. The target for generator training is always one. We want the discriminator to
give us the gradients that would have made a fake image look like a real one.
14. Just to be sure, we set the discriminator to be non-trainable, so that we can
not break anything by accident.
15. We train the GAN model. We feed in a batch of random latent vectors and
train the generator part of the GAN so that the discriminator part will classify
the generated images as real.
16. We save the losses from training.
In the following figure, you can see some of the generated MNIST characters:
Most of these characters look like identifiable numbers, although some, such as those
in the bottom left and right, seem a bit off.
[ 261 ]
Using Generative Models
The code that we wrote and explored is now outputted in the following chart,
showing us the Discriminitive and Generative loss of an increasing number of
Epochs.
Note that the loss in GAN training is not interpretable as it is for supervised learning.
The loss of a GAN will not decrease even as the GAN makes progress.
The loss of a generator and discriminator is dependent on how well the other model
does. If the generator gets better at fooling the discriminator, then the discriminator
loss will stay high. If one of the losses goes to zero, it means that the other model lost
the race and cannot fool or properly discriminate the other model anymore.
This is one of the things that makes GAN training so hard: GANs don't converge to
a low loss solution; they converge to an equilibrium in which the generator fools the
discriminator not all the time, but many times. That equilibrium is not always stable.
Part of the reason so much noise is added to labels and the networks themselves is
that it increases the stability of the equilibrium.
As GANs are unstable and difficult, yet useful, a number of tricks has been
developed over time that makes GAN training more stable. Knowing these tricks
can help you with your GAN building process and save you countless hours, even
though there is often no theoretical reason for why these tricks work.
[ 262 ]
Chapter 6
Research has shown that the latent space of GANs often has some surprising
features, such as "smile vectors," which arrange face images according to the width
of the person's smile. Researchers have also shown that GANs can be used for latent
space algebra, where adding the latent representation of different objects creates
realistic, new objects. Yet, research on the latent space of GANs is still in its infancy
and drawing conclusions about the world from its latent space representations is an
active field of research.
• Normalize the inputs: GANs don't work well with extreme values, so make
sure you always have normalized inputs between -1 and 1. This is also the
reason why you should use the tanh function as your generator output.
• Don't use the theoretical correct loss function: If you read papers on
GANs, you will find that they give the generator optimization goal as
the following formula:
max log D
[ 263 ]
Using Generative Models
[ 264 ]
Chapter 6
If the generator goes completely off track and produces only zeros, for
instance, you will be able to see it before spending days of GPU time
on training that will go nowhere.
• Don't balance loss via statistics: Keeping the balance between the generator
and discriminator is a delicate task. Many practitioners, therefore, try
to help the balance by training either the generator or discriminator a bit
more depending on statistics. Usually, that does not work. GANs are very
counterintuitive and trying to help them with an intuitive approach usually
makes matters worse. That is not to say there are no ways to help out GAN
equilibriums, but the help should stem from a principled approach, such as
"train the generator while the generator loss is above X."
• If you have labels, use them: A slightly more sophisticated version of a GAN
discriminator can not only classify data as real or fake but also classify the class
of the data. In the MNIST case, the discriminator would have 11 outputs: an
output for the 10 real numbers as well as an output for a fake. This allows us to
create a GAN that can show more specific images. This is useful in the domain
of semi-supervised learning, which we will cover in the next section.
• Add noise to inputs, reduce it over time: Noise adds stability to GAN
training so it comes as no surprise that noisy inputs can help, especially in
the early, unstable phases of training a GAN. Later, however, it can obfuscate
too much and keep the GAN from generating realistic images. So, we should
reduce the noise applied to inputs over time.
• Use dropouts in G in both the train and test phases: Some researchers find
that using dropout on inference time leads to better results for the generated
data. Why that is the case is still an open question.
• Historical averaging: GANs tend to "oscillate," with their weights moving
rapidly around a mean during training. Historical averaging penalizes
weights that are too far away from their historical average and reduces
oscillation. It, therefore, increases the stability of GAN training.
• Replay buffers: Replay buffers keep a number of older generated images
so they can be reused for training the discriminator. This has a similar effect
as historical averaging, reducing oscillation and increasing stability. It also
reduces the correlation and the test data.
• Target networks: Another "anti-oscillation" trick is to use target networks.
That is, to create copies of both the generator and discriminator, and
then train the generator with a frozen copy of the discriminator and train
the discriminator with a frozen copy of the generator.
• Entropy regularization: Entropy regularization means rewarding the network
for outputting more different values. This can prevent the generator network
from settling on a few things to produce, say, only the number seven. It is
a regularization method as it prevents overfitting.
[ 265 ]
Using Generative Models
• Use dropout or noise layers: Noise is good for GANs. Keras not only
features dropout layers, but it also features a number of noise layers that
add different kinds of noise to activations in a network. You can read the
documentation of these layers to see whether they are helpful for your
specific GAN application: https://keras.io/layers/noise/.
Supervised learning uses data with labels. An example is the image classifier we
built in Chapter 3, Utilizing Computer Vision, or most of the other models that we've
built in this book.
MTurk is an Amazon website that offers, according to its own site, "Human
intelligence through an API." In practice, this means that companies and researchers
post simple jobs such as filling out a survey or classifying an image and people
all over the world perform these tasks for a few cents per task. For an AI to learn,
humans need to provide labeled data. If the task is large scale, then many companies
will hire MTurk users in order to let humans do the labeling. If it is a small task, you
will often find the company's own staff labeling the data.
[ 266 ]
Chapter 6
Surprisingly little thought goes into what these humans label. Not all labels are equally
useful. The following diagram shows a linear classifier. As you can see, the frontier
point, which is close to the frontier between the two classes, determines where the
decision boundary is, while the points further in the back are not as relevant:
As such, the frontier points are more valuable than points further away from the
decision boundary. You can train on less data by doing the following:
This process of labeling data is much more efficient than just randomly labeling
data and can accelerate your efforts quite drastically.
[ 267 ]
Using Generative Models
Another way is to once again label a few images and pretrain a weak model. At the
point where the labeling takes place, the computer shows the labeler the data as
well as a label. The labeler only has to decide whether this label is correct. This can
be done easily with hotkeys and the time it takes to label a single item goes down
dramatically. If the label was wrong, the label interface can either bring up a list of
possible options, sorted by the probability the model assigned to them, or just put
the item back on the stack and display the next most likely label the next time.
A great implementation of this technique is "Prodigy," a labeling tool by the
company that makes spaCy, which we learned about in Chapter 5, Parsing Textual
Data with Natural Language Processing, we can see an example of the Prodigy tool
in the following screenshot:
Prodigy is a labeling tool that leverages machines, you can find more out about it by
reading its official documentation here: https://prodi.gy/.
The first output of the discriminator only classifies data as real or fake, just as it
did for the GAN previously. The second output classifies the data by its class, for
example, the digit an image represents, or an extra "is fake" class. In the MNIST
example, the classifying output would have 11 classes, 10 digits plus the "is fake"
class. The trick is that the generator is one model and only the output, that is, the last
layer, is different. This forces the "real or not" classification to share weights with the
"which digit" classifier.
The idea is that to determine whether an image is real or fake, the classifier would
have to figure out whether it can classify this image into one class. If it can, the image
is probably real. This approach, called semi-supervised generative adversarial
network (SGAN), has been shown to generate more realistic data and deliver better
results on limited data than standard supervised learning. Of course, GANs can be
applied to more than just images.
In the next section, we will apply them to our fraud detection task.
SGAN scheme
[ 269 ]
Using Generative Models
We will train this model on fewer than 1,000 transactions and still get a decent fraud
detector.
Note: You can find the code for the SGAN on Kaggle under this link:
https://www.kaggle.com/jannesklaas/semi-supervised-
gan-for-fraud-detection/code.
In this case, our data has 29 dimensions. We set our latent vectors to have
10 dimensions:
latent_dim=10
data_dim=29
To use the generator model better, we wrap the model we created into a functional
API model that maps the noise vector to a generated transaction record. Since most
GAN literature is about images, and "transaction record" is a bit of a mouthful,
therefore we just name our transaction records as "images":
noise = Input(shape=(latent_dim,))
img = model(noise)
Just as we did with the generator, we build the discriminator in the sequential API.
The discriminator has two output: one for the classes, and one for fake or not fake.
We first only construct the base of the model:
model = Sequential()
model.add(Dense(31,input_dim=data_dim))
model.add(LeakyReLU(alpha=0.2))
model.add(BatchNormalization(momentum=0.8))
model.add(Dropout(0.25))
model.add(Dense(16,input_dim=data_dim))
model.add(LeakyReLU(alpha=0.2))
[ 270 ]
Chapter 6
Now we'll map the input of the discriminator to its two heads using the
functional API:
img = Input(shape=(data_dim,)) #1
features = model(img) #2
valid = Dense(1, activation="sigmoid")(features) #3
label = Dense(num_classes+1, activation="softmax")(features) #4
Let's take a minute to look at the five key aspects of the preceding code:
To compile the discriminator with two heads, we need to use a few advanced model-
compiling tricks:
optimizer = Adam(0.0002, 0.5) #1
discriminator.compile(loss=['binary_crossentropy',
'categorical_crossentropy'], #2
loss_weights=[0.5, 0.5], #3
optimizer=optimizer, #4
metrics=['accuracy']) #5
[ 271 ]
Using Generative Models
Again, looking at the code, we can see the following key points:
For training, we define a train function, which handles all the training for us:
def train(X_train,y_train,
X_test,y_test,
generator,discriminator,
combined,
num_classes,
epochs,
batch_size=128):
f1_progress = [] #1
half_batch = int(batch_size / 2) #2
cw1 = {0: 1, 1: 1} #3
cw2 = {i: num_classes / half_batch for i in
range(num_classes)}
cw2[num_classes] = 1 / half_batch
[ 272 ]
Chapter 6
gen_imgs = generator.predict(noise)
labels = to_categorical(y_train[idx],
num_classes=num_classes+1) #7
if epoch % 10 == 0: #15
_,y_pred = discriminator.predict(X_test,
batch_size=batch_size)
y_pred = np.argmax(y_pred[:,:-1],axis=1)
f1 = f1_score(y_test,y_pred)
print('Epoch: {}, F1: {:.5f}'.format(epoch,f1))
f1_progress.append(f1)
return f1_progress
[ 273 ]
Using Generative Models
Take a minute to pause: that was a very long and complicated function to code.
Before we summarize this chapter, let's look at the 15 key elements of that code:
[ 274 ]
Chapter 6
Now that we have everything set up, we are going to train our SGAN for 5,000
epochs. This will take about 5 minutes on a GPU but could take much longer if
you do not have a GPU:
f1_p = train(X_res,y_res,
X_test,y_test,
generator,discriminator,
combined,
num_classes=2,
epochs=5000,
batch_size=128)
Finally, we plot the F1 score of our semi-supervised fraud classifier over time:
fig = plt.figure(figsize=(10,7))
plt.plot(f1_p)
plt.xlabel('10 Epochs')
plt.ylabel('F1 Score Validation')
SGAN progress
[ 275 ]
Using Generative Models
As you can see, the model learns pretty quickly at first, but then collapses with its F1
score going to zero. This is a textbook example of a collapsing GAN. As mentioned
previously, GANs are unstable. If the delicate balance between the generator and
discriminator breaks, performance quickly deteriorates.
Making GANs more stable is an active area of research. So far, many practitioners
will just attempt to try multiple runs with different hyperparameters and random
seeds in a hope to get lucky. Another popular method is just to save the model every
couple of epochs. The model seems to be a pretty decent fraud detector at around
epoch 150 despite being trained on fewer than 1,000 transactions.
Exercises
To get more comfortable with generative models, try your hand at these exercises:
1. Create an SGAN in order to train an MNIST image classifier. How few
images can you use to achieve over 90% classification accuracy?
2. Using LSTMs, you can build an autoencoder for stock price movements.
Using a dataset such as the DJIA stock prices, build an autoencoder that
encodes stock movements. Then visualize what happens to the outputs as
you move through the latent space. You can find the dataset here: https://
www.kaggle.com/szrlee/stock-time-series-20050101-to-20171231.
Summary
In this chapter, you have learned about the two most important types of generative
models: autoencoders and GANs. We first developed an autoencoder for MNIST
images. We then used a similar architecture to encode credit card data and detect
fraud. Afterward, we expanded the autoencoder to a VAE. This allowed us to learn
distributions of encodings and generate new data that we could use for training.
Afterward, we learned about GANs, again first in the context of MNIST images and
then in the context of credit card fraud. We used an SGAN to reduce the amount of
data we needed to train our fraud detector. We used model outputs to reduce the
amount of labeling necessary through active learning and smarter labeling interfaces.
We've also discussed and learned about latent spaces and the use they have for
financial analysis. We saw the t-SNE algorithm and how it can be used to visualize
higher dimensional (latent) data. You also got a first impression of how machine
learning can solve game-theoretic optimization problems. GANs solve a minimax
problem, which is frequent in economics and finance.
In the next chapter, we will deep dive into exactly that type of optimization as we
cover reinforcement learning.
[ 276 ]
Reinforcement Learning
for Financial Markets
Humans don't learn from millions of labeled examples. Instead, we often learn from
positive or negative experiences that we associate with our actions. Children that
touch a hot stove once will never touch it again. Learning from experiences and the
associated rewards or punishments is the core idea behind reinforcement learning
(RL). RL allows us to learn sophisticated decision-making rules while having no
data at all. Through this approach, several high-profile breakthroughs occurred
in AI, such as AlphaGo, which beat the world Go champion in 2016.
[ 277 ]
Reinforcement Learning for Financial Markets
This chapter will start with an intuitive introduction to RL using a simple "catch
the fruit" game. We will then dive into the underlying theory before covering more
advanced RL applications. The examples in this chapter rely on visualizations that
are not easily rendered in Kaggle kernels. In order to simplify them, the example
algorithms are also not optimized for GPU usage. It is, therefore, best to run these
examples on your local machine.
The algorithms in this chapter run relatively quickly, so you won't have to wait too
long for them to run. The chapter code was written on a Mid-2012 MacBook Pro, and
no example took longer than 20 minutes to run on that machine. Of course, you can
also run the code on Kaggle, however the visualizations will not work there.
The goal here is to let the computer play Catch by itself. We will be using a simplified
version in this example in order to make the task easier:
[ 278 ]
Chapter 7
While playing Catch, the player decides between three possible actions. They can
move the basket to the left, to the right, or make it stay put.
The basis for this decision is the current state of the game; in other words, the
positions of the falling fruit and of the basket. Our goal is to create a model that,
given the content of the game screen, chooses the action that leads to the highest
score possible. This task can be seen as a simple classification problem. We could ask
expert human players to play the game multiple times and record their actions. Then,
we could train a model to choose the "correct" action that mirrors the expert players.
This is not how humans learn, however. Humans can learn a game such as Catch
by themselves, without guidance. This is very useful, because imagine if you had to
hire a bunch of experts to perform a task thousands of times every time you wanted
to learn something as simple as Catch: it would be expensive and slow.
In reinforcement learning, the model trains from experience, rather than labeled data.
Instead of providing the model with the correct actions, we provide it with rewards
and punishments. The model receives information about the current state of the
environment, for example, the computer game screen. It then outputs an action, such
as a joystick movement. The environment reacts to this action and provides the next
state, along with any rewards:
RL scheme
The model then learns to find actions that lead to maximum rewards. There are
many ways this can work in practice. Right now, we are going to look at Q-learning.
Q-learning made a splash when it was used to train a computer to play Atari video
games. Today, it is still a relevant concept. Most modern RL algorithms are based on
some adaptation of Q-learning.
As a result, there will be some reward, r, and a new state, s ′ . The problem with
both Catch and chess is that the rewards do not appear immediately after the action.
In Catch, you only earn rewards when the fruits hit the basket or fall on the floor,
and in chess, you only earn a reward when you win or lose the game. This means
that rewards are sparsely distributed. Most of the time, r will be zero. When there
is a reward, it is not always a result of the action taken immediately before. Some
action taken long before might have caused the victory. Figuring out which action
is responsible for the reward is often referred to as the credit assignment problem.
Because rewards are delayed, good chess players do not choose their plays only
by the immediate reward. Instead, they choose the expected future reward.
For example, they do not only think about whether they can eliminate an opponent's
figure in the next move, they also consider how taking a specific action now will
help them in the long run. In Q-learning, we choose our action based on the highest
expected future reward. We use a Q-function to calculate this. This is a mathematical
function that takes two arguments: the current state of the game, and a given
action. We can write this as Q(state, action).
While in state s, we estimate the future reward for each possible action, a. We
assume that after we have taken action a and moved to the next state, s ′, everything
works out perfectly. The expected future reward, q(s,a), for a given state and action
is calculated as the immediate reward, plus the expected future reward thereafter,
Q ( s ′, a ′) . We assume the next action, a ′ , is optimal. Because there is uncertainty
about the future, we discount Q ( s ′, a ′) by the factor gamma, γ . We, therefore,
arrive at an expected reward of this:
Q ( s, a ) = r + γ ∗ max Q ( s ′, a ′)
Good chess players are very good at estimating future rewards in their head. In other
words, their Q-function, Q(s,a), is very precise.
[ 280 ]
Chapter 7
These experiences are our training data. We can frame the problem of estimating
Q(s,a) as a regression problem. To solve this, we can use a neural network. Given an
input vector consisting of s and a, the neural network is supposed to predict the value
of Q(s,a) equal to the target: r + γ ∗ maxQ ( s ′, a ′) . If we are good at predicting Q(s,a) for
different states s and actions a, we will have a good approximation of the Q-function.
Given a batch of experiences, < s, a, r , s ′ > , the training process then looks as follows:
1. For each possible action, a ′ , (left, right, stay), predict the expected future
reward, Q ( s ′, a ′) , using the neural network.
2. Choose the highest value of the three predictions as the max, Q ( s ′, a ′) .
3. Calculate r + γ ∗ maxQ ( s ′, a ′) . This is the target value for the neural network.
4. Train the neural network using a loss function. This is a function that
calculates how near or far the predicted value is from the target value. Here,
2
we will use 0.5 ∗ ( predictedQ ( s, a)− target ) as the loss function. Effectively, we want
to minimize the squared error between prediction and target. The factor of
0.5 is just there to make the gradient nicer.
During gameplay, all the experiences are stored in a replay memory. This acts like
a simple buffer in which we store < s, a, r , s ′ > pairs. The ExperienceReplay class
also handles preparing the data for training.
[ 281 ]
Reinforcement Learning for Financial Markets
self.discount = discount
if game_over: #13
targets[i, action_t] = reward_t
else:
targets[i, action_t] = reward_t + self.discount * Q_sa
return inputs, targets
Let's pause for a second and break down the code that we've just created:
[ 282 ]
Chapter 7
2. To initialize a replay object, we need to let it know how large its buffer
should be and what the discount rate, γ , is. The replay memory itself is a list
of lists following this scheme:
[...
[experience, game_over]
[experience, game_over]
...]
[ 283 ]
Reinforcement Learning for Financial Markets
As this is a regression problem, the final layer has no activation function, and the
loss is a mean squared error loss. We optimize the network using stochastic gradient
descent without momentum or any other bells and whistles:
model = Sequential()
model.add(Dense(100, input_shape=(grid_size**2,),
activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(num_actions))
model.compile(optimizer='sgd', loss='mse')
The same goes for Q-learning. By always choosing the best option, you might miss
out on some unexplored paths. To avoid this, the learner will sometimes choose
a random option, and not necessarily the best one.
win_hist = []
for e in range(epochs): #2
loss = 0.
env.reset()
game_over = False
input_t = env.observe()
[ 284 ]
Chapter 7
loss += batch_loss
win_hist.append(win_cnt)
return win_hist
Before we go further, again let's break down the code so we can see what we're
doing:
[ 285 ]
Reinforcement Learning for Financial Markets
The following graph shows the rolling mean of successful games. After about
2,000 epochs of training, the neural network should be quite good at playing Catch:
Looking at the preceding graph, it's safe to say that you have now successfully
created your first reinforcement learning system, as after 5000 epochs the average of
victories per games is between 90% and 100%. In the next section, we will explore the
theoretical foundations of reinforcement learning and discover how the same system
that learns to play catch can learn to route orders in the futures market.
[ 286 ]
Chapter 7
A Markov model describes a stochastic process with different states in which the
probability of ending up in a specific state is purely dependent on the state one
is currently in. In the following diagram, you can see a simple Markov model
describing recommendations given for a stock:
As you can see, there are three states in this model, BUY, HOLD, and SELL. For
every two states, there is a transition probability. For example, the probability
that a state gets a BUY recommendation if it had a HOLD recommendation in the
previous round is described by p ( BUY | HOLD) , which is equal to 0.5. There is a 50%
chance that a stock that is currently in HOLD will move to BUY in the next round.
[ 287 ]
Reinforcement Learning for Financial Markets
States are associated with rewards. If you own stock, and that stock has a BUY
recommendation, the stock will go up, and you will earn a reward of 1. If the stock
has a sell recommendation, you will gain a negative reward, or punishment, of -1.
Note: In some textbooks, the rewards are associated with state transitions
and not states themselves. It turns out to be mathematically equivalent,
and for the ease of notation, we are associating the rewards with states
here.
V π ( s ) = E π [ Rt | st = s ]
The expected return is the reward gained immediately plus the discounted future
rewards:
∞
Rt = rt +1 + γ ∗ rt +2 + γ 2 ∗ rt +3 + γ 3 ∗ rt +3 = ∑ γ k ∗ rt +k +1
k =0
The other value function frequently used in RL is the function Q(s,a), which we have
already seen in the previous section. Q describes the expected return of taking action
a in state s if policy π is followed:
Q π ( s, a ) = E [ Rt | st = s, at = a ]
[ 288 ]
Chapter 7
Note: We use the expected value since our environment and our actions
are stochastic. We cannot say for certain that we will land in a specific
state; we can only give a probability.
Q and V describe the same thing. If we find ourselves in a certain state, what should
we do? V gives recommendations regarding which state we should seek, and Q gives
advice on which action we should take. Of course, V implicitly assumes we have
to take some action and Q assumes that the result of our actions is landing in some
state. In fact, both Q and V are derived from the so-called Bellman equation, which
brings us back to the Markov model from the beginning of this section.
If you assume that the environment you operate in can be described as a Markov
model, you would really want to know two things. First, you would want to find
out the state transition probabilities. If you are in state s, what is the chance Pssa′ ends
in state s ′ if you take action a? Mathematically, it is as follows:
Pssa′ = Pr ( st +1 = s ′ | st = s, at = a)
Equally, you would be interested in the expected reward Rssa ′ of being in state s,
taking action a and ending up in state s ′ :
Rssa ′ = E rt +1 | st = s, st +1 = s ′, at = a
With this in mind, we can now derive the two Bellman equations for Q and V. First,
we rewrite the equation describing V to contain the actual formula for Rt :
∞
V π = E π ∑ γ k ∗ rt +k +1 | st = s
k =0
∞
V π ( s ) = E π rt +1 + γ ∗ ∑ γ k ∗ rt +k +2 | st = s
k =0
The first part of our expectation is the expected reward we directly receive from
being in state s and the following policy, π :
E π rt +1 | st = s = ∑ π ( s, a )∑ Pssa′ Rssa ′
a s′
[ 289 ]
Reinforcement Learning for Financial Markets
The preceding equation shows a nested sum. First, we sum over all actions, a,
weighted by their probability of occurrence under policy π . For each action, we
then sum over the distribution of rewards from the transition from state s to the
next state, s ′ , after action a, weighted by the probability of this transition occurring
following the transition probability .
The expected discounted value of the future rewards after state s is the discounted
expected future value of all states, s ′ , weighted by their probability of occurrence, ,
and the probability of action a being taken following policy π .
This formula is quite a mouthful, but it gives us a glimpse into the recursive nature
of the value function. If we now replace the expectation in our value function, it
becomes clearer:
∞
V π ( s ) = ∑ π ( s, a ) ∑ Pssa′ Rssa ′ + γ E π ∑ γ k rt +k +2 st +1 = s ′
k =0
a s′
The inner expectation represents the value function for the next step, s ′ ! That means
we can replace the expectation with the value function, V ( s ′) :
Congratulations, you have just derived the Bellman equation! For now, pause
and take a second to ponder and make sure you really understand the mechanics
behind these equations. The core idea is that the value of a state can be expressed
as the value of other states. For a long time, the go-to approach to optimizing the
Bellman equation was to build a model of the underlying Markov model and its
state transition and reward probabilities.
[ 290 ]
Chapter 7
However, the recursive structure calls for a technique called dynamic programming.
The idea behind dynamic programming is to solve easier sub-problems. You've
already seen this in action in the Catch example. There, we used a neural network
to estimate Q π (s ′, a ′) , except for states that ended the game. For these games, finding
the reward associated with the state is easy: it is the final reward received at the
end of the game. It was these states for which the neural network first developed an
accurate estimate of the function Q. From there, it could then go backward and learn
the values of states that were further away from the end of the game. There are more
possible applications of this dynamic programming and model-free approach to
reinforcement learning.
Before we jump into the different kinds of systems that can be built using this
theoretical foundation, we will pay a brief visit to the applications of the Bellman
equation in economics. Readers who are familiar with the work discussed here will
find reference points that they can use to develop a deeper understanding of Bellman
equations. Readers unfamiliar with these works will find inspiration for further
reading and applications for techniques discussed throughout this chapter.
[ 291 ]
Reinforcement Learning for Financial Markets
A( s, a ) = Qπ ( s, a ) −V p i ( s )
J (π ) = E p s0 V ( s0 )
Now, to compute the gradient of the policy, we have to do two steps, which are
shown inside the expectation in the policy gradient formula:
[ 292 ]
Chapter 7
First, we have to calculate the advantage of a given action, a, with A(s,a). Then we
have to calculate the derivative of the weights of the neural network, ∇θ , with respect
to increasing the probability, π (a | s ) , that a is picked under policy π .
For actions with a positive advantage, A(s,a), we follow the gradient that would
make a more likely. For actions with a negative advantage, we go in the exact
opposite direction. The expectation says that we are doing this for all states and
all actions. In practice, we manually multiply the advantage of actions with their
increased likelihood gradients.
One thing left for us to look at is how we compute the advantage. The value of taking
an action is the reward earned directly as a result of taking the action, as well as the
value of the state we find ourselves in after taking that action:
Q ( s, a ) = r + γV ( s ′)
A( s, a ) = Q ( s, a ) −V ( s ) = r + γV ( s ′) −V ( s )
A2C scheme
If you are training an agent that operates on high-dimensional image data, for
instance, the value function and the policy head then both need to learn how to
interpret the image. Sharing weights would help master the common task. If you are
training on lower dimensional data, it might make more sense to not share weights.
[ 293 ]
Reinforcement Learning for Financial Markets
Since both A3C and A2C work by the same principles, and the implementation of
parallel gameplay introduces some complexity that obfuscates the actual algorithm,
we will just stick with A2C in the following examples.
Learning to balance
In this section, we will train an A2C model to swing up and balance a pendulum:
Pendulum gym
[ 294 ]
Chapter 7
The pendulum environment is part of the OpenAI Gym, a suite of games made to train
reinforcement learning algorithms. You can install it via the command line as follows:
pip install gym
import numpy as np #2
There are quite a few new imports, so let's walk through them one by one:
[ 295 ]
Reinforcement Learning for Financial Markets
Now it is time to build the agent. The following methods all form the A2CAgent class:
def __init__(self, state_size, action_size):
self.state_size = state_size #1
self.action_size = action_size
self.value_size = 1
self.exp_replay = deque(maxlen=2000) #2
self.actor_lr = 0.0001 #3
self.critic_lr = 0.001
self.discount_factor = .9
self.optimize_actor = self.actor_optimizer() #5
self.optimize_critic = self.critic_optimizer()
1. First, we need to define some game-related variables. The state space size and
the action space size are given by the game. Pendulum states consist of three
variables dependent on the angle of the pendulum. A state consists of the
sine of theta, the cosine of theta, and the angular velocity. The value of a state
is just a single scalar.
2. Next, we set up our experience replay buffer, which can save at maximum
2,000 states. Larger RL experiments have much larger replay buffers
(often around 5 million experiences), but for this task 2,000 will do.
3. As we are training a neural network, we need to set some hyperparameters.
Even if the actor and critic share weights, it turns out that the actor learning
rate should usually be lower than the critic learning rate. This is because the
policy gradient we train the actor on is more volatile. We also need to set the
discount rate, γ . Remember that the discount rate in reinforcement learning
is applied differently than it is usually in finance. In finance, we discount
by dividing future values by one plus the discount factor. In reinforcement
learning, we multiply with the discount rate. Therefore, a higher discount
factor, γ , means that future values are less discounted.
4. To actually build the model, we define a separate method, which we will
discuss next.
[ 296 ]
Chapter 7
5. The optimizers for actor and critic are custom optimizers. To define these, we
also create a separate function. The optimizers themselves are functions that
can be called at training time:
def build_model(self):
actor_input = Dense(30, #2
activation='relu',
kernel_initializer='he_uniform')(state)
mu_0 = Dense(self.action_size, #3
activation='tanh',
kernel_initializer='he_uniform')(actor_input)
mu = Lambda(lambda x: x * 2)(mu_0) #4
sigma_0 = Dense(self.action_size, #5
activation='softplus',
kernel_initializer='he_uniform')(actor_input)
critic_input = Dense(30, #7
activation='relu',
kernel_initializer='he_uniform')(state)
actor._make_predict_function() #11
critic._make_predict_function()
actor.summary() #12
critic.summary()
[ 297 ]
Reinforcement Learning for Financial Markets
The preceding function sets up the Keras model. It is quite complicated, so let's go
through it:
1. As we are using the functional API, we have to define an input layer that we
can use to feed the state to the actor and critic.
2. The actor has a hidden first layer as an input to the actor value function.
It has 30 hidden units and a relu activation function. It is initialized by
an he_uniform initializer. This initializer is only slightly different from the
default glorot_uniform initializer. The he_uniform initializer draws from
a uniform distribution with the limits ± 6 i , where i is the input dimension.
The default glorot uniform samples from a uniform distribution with the
limits ± 6 (i + 0) , with o being the output dimensionality. The difference between
the two is rather small, but as it turns out, the he_uniform initializer works
better for learning the value function and policy.
3. The action space of the pendulum ranges from -2 to 2. We use a regular tanh
activation, which ranges from -1 to 1 first and corrects the scaling later.
4. To correct the scaling of the action space, we now multiply the outputs of the
tanh function by two. Using the Lambda layer, we can define such a function
manually in the computational graph.
5. The standard deviation should not be negative. The softplus activation
works in principle just like relu, but with a soft edge:
[ 298 ]
Chapter 7
6. To make sure the standard deviation is not zero, we add a tiny constant to
it. Again we use the Lambda layer for this task. This also ensures that the
gradients get calculated correctly, as the model is aware of the constant
added.
7. The critic also has a hidden layer to calculate its value function.
8. The value of a state is just a single scalar that can have any value. The value
head thus only has one output and a linear, that is: the default, activation
function.
9. We define the actor to map from a state to a policy as expressed by the mean,
µ , and standard deviation, σ .
10. We define the critic to map from a state to a value of that state.
11. While it is not strictly required for A2C, if we want to use our agents for an
asynchronous, A3C approach, then we need to make the predict function
threading safe. Keras loads the model on a GPU the first time you call
predict(). If that happens from multiple threads, things can break. _make_
predict_function() makes sure the model is already loaded on a GPU
or CPU and is ready to predict, even from multiple threads.
12. For debugging purposes, we print the summaries of our models.
13. Finally, we return the models.
Now we have to create the optimizer for the actor. The actor uses a custom optimizer
that optimizes it along the policy gradient. Before we define the optimizer; however,
we need to look at the last piece of the policy gradient. Remember how the policy
gradient was dependent on the gradient of the weights, ∇θ π (a | s) , that would make
action a more likely? Keras can calculate this derivative for us, but we need to
provide Keras with the value of policy π .
2πσ 2
In this term, π stands for the constant, 3.14…, not for the policy. Later, we only
need to take the logarithm of this probability density function. Why the logarithm?
Because taking the logarithm results in a smoother gradient. Maximizing the log of
a probability means maximizing the probability, so we can just use the "log trick,"
as it is called, to improve learning.
[ 299 ]
Reinforcement Learning for Financial Markets
The value of policy π is the advantage of each action a, times the log probability
of this action occurring as expressed by the probability density function, f.
The following function optimizes our actor model. Let's go through the optimization
procedure:
def actor_optimizer(self):
action = K.placeholder(shape=(None, 1)) #1
advantages = K.placeholder(shape=(None, 1))
optimizer = Adam(lr=self.actor_lr) #9
updates = optimizer.get_updates(self.actor.trainable_weights,
[], actor_loss) #10
1. First, we need to set up some placeholders for the action taken and the
advantage of that action. We will fill in these placeholders when we call
the optimizer.
2. We get the outputs of the actor model. These are tensors that we can plug
into our optimizer. Optimization of these tensors will be backpropagated
and optimizes the whole model.
3. Now we set up the probability density function. This step can look a bit
intimidating, but if you look closely, it is the same probability density
function we defined previously.
[ 300 ]
Chapter 7
4. Now we apply the log trick. To ensure that we don't accidentally take
the logarithm of zero, we add a tiny constant, epsilon.
5. The value of our policy is now the probability of action a times the
probability of this action occurring.
6. To reward the model for a probabilistic policy, we add an entropy term.
The entropy is calculated with the following term:
For the critic, we also need to create a custom optimizer. The loss for the critic is the
mean squared error between the predicted value and the reward plus the predicted
value of the next state:
def critic_optimizer(self):
discounted_reward = K.placeholder(shape=(None, 1)) #1
value = self.critic.output
[ 301 ]
Reinforcement Learning for Financial Markets
optimizer = Adam(lr=self.critic_lr) #3
updates = optimizer.get_updates(self.critic.trainable_weights,
[], loss)
return train
For our agent to take actions, we need to define a method that produces actions
from a state:
def get_action(self, state):
state = np.reshape(state, [1, self.state_size]) #1
mu, sigma_sq = self.actor.predict(state) #2
epsilon = np.random.randn(self.action_size) #3
action = mu + np.sqrt(sigma_sq) * epsilon #4
action = np.clip(action, -2, 2) #5
return action
With this function, our actor can now act. Let's go through it:
1. First, we reshape the state to make sure it has the shape the model expects.
2. We predict the means and variance, σ 2 , for this action from the model.
[ 302 ]
Chapter 7
At last, we need to train the model. The train_model function will train the model
after receiving one new experience:
def train_model(self, state, action, reward, next_state, done):
self.exp_replay.append((state, action, reward,
next_state, done)) #1
value = self.critic.predict(state)[0] #4
next_value = self.critic.predict(next_state)[0]
if done: #5
advantages[0] = reward - value
target[0][0] = reward
else:
advantages[0] = reward + self.discount_factor *
(next_value) - value
target[0][0] = reward + self.discount_factor * next_value
[ 303 ]
Reinforcement Learning for Financial Markets
3. We set up placeholders for the advantages and targets. We will fill them
at step 5.
4. We predict the values for state s and s ′ .
5. If the game ended after the current state, s, the advantage is the reward
we earned minus the value we assigned to the state, and the target for
the value function is just the reward we earned. If the game did not end
after this state, the advantage is the reward earned plus the discounted
value of the next state minus the value of this state. The target, in that case,
is the reward earned plus the discounted value of the next state.
6. Knowing the advantage, the action taken, and the value target, we can
optimize both the actor and critic with the optimizers we created earlier.
And that is it; our A2CAgent class is done. Now it is time to use it. We define a
run_experiment function. This function plays the game for a number of episodes. It
is useful to first train a new agent without rendering, because training takes around
600 to 700 games until the agent does well. With your trained agent, you can then
watch the gameplay:
def run_experiment(render=False, agent=None, epochs = 3000):
env = gym.make('Pendulum-v0') #1
state_size = env.observation_space.shape[0] #2
action_size = env.action_space.shape[0]
if agent = None: #3
agent = A2CAgent(state_size, action_size)
scores = [] #4
for e in range(epochs): #5
done = False #6
score = 0
state = env.reset()
state = np.reshape(state, [1, state_size])
action = agent.get_action(state) #9
next_state, reward, done, info = env.step(action) #10
reward /= 10 #11
next_state = np.reshape(next_state,
[ 304 ]
Chapter 7
if done: #16
scores.append(score)
print("episode:", e, " score:", score)
[ 305 ]
Reinforcement Learning for Financial Markets
13. Now we train the agent on a new experience. As you have seen, the agent
will store the experience in its replay buffer and draw a random old
experience to train from.
14. We increase the overall reward to track the rewards earned during one game.
15. We set the new state to be the current state to prepare for the next frame
of the game.
16. If the game is over, we track and print out the game score.
17. The agent usually does pretty well after 700 epochs. We declare the game
solved if the average reward over the last 20 games was better than -20.
If that is the case, we will exit the function and return the trained agent
together with its scores.
Learning to trade
Reinforcement learning algorithms are largely developed in games and simulations
where a failing algorithm won't cause any damage. However, once developed, an
algorithm can be adapted to other, more serious tasks. To demonstrate this ability,
we are now going to create an A2C agent that learns how to balance a portfolio of
stocks within a large universe of stocks.
The allocation vector describes the share of assets the agent wants to allocate
on one stock. A negative allocation means the agent is short trading the stock. For
simplicity's sake, transaction costs and slippage are not added to the environment.
It would not be too difficult to add them, however.
Tip: The full implementation of the environment and agent can be found
at https://www.kaggle.com/jannesklaas/a2c-stock-trading.
[ 306 ]
Chapter 7
return self.data[0,:-1,:] #5
def step(self,allocation): #6
ret = np.sum(allocation * self.data[self.pos,-1,:]) #7
self.returns.append(ret) #8
mean = 0 #9
std = 1
if len(self.returns) >= 20: #10
mean = np.mean(self.returns[-20:])
std = np.std(self.returns[-20:]) + 0.0001
df = pd.concat(frames,axis=1,ignore_index=False) #16
[ 307 ]
Reinforcement Learning for Financial Markets
df = df.pct_change()
df = df.fillna(0)
batch = df.values
episodes = [] #17
for i in range(batch.shape[0] - 101):
eps = batch[i:i+101]
episodes.append(eps)
data = np.stack(episodes)
assert len(data.shape) == 3
assert data.shape[-1] == 100
return data
Our trade environment is somewhat similar to the pendulum environment. Let's see
how we set it up:
1. We load data for our universe.
2. Since we are stepping through the data where each day is a step, we need
to keep track of our position in time.
3. We need to know when the game ends, so we need to know how much data
we have.
4. To keep track of returns over time, we set up an empty array.
5. The initial state is the data for the first episode, until the last element, which
is the return of the next day for all 100 stocks in the universe.
6. At each step, the agent needs to provide an allocation to the environment.
The reward the agent receives is the sharpe ratio, the ratio between the mean
and standard deviation of returns, over the last 20 days. You could modify
the reward function to, for example, include transaction costs or slippage.
If you do want to do this, then refer to the section on reward shaping later
in this chapter.
7. The return on the next day is the last element of the episode data.
8. To calculate the Sharpe ratio, we need to keep track of past returns.
9. If we do not have 20 returns yet, the mean and standard deviation of returns
will be zero and one respectively.
10. If we do have enough data, we calculate the mean and standard deviation
of the last 20 elements in our return tracker. We add a tiny constant to the
standard deviation to avoid division by zero.
11. We can now calculate the Sharpe ratio, which will provide the reward for
the agent.
12. If the game is over, the environment will return no next state, the reward,
and an indicator that the game is over, as well as an empty information
dictionary in order to stick to the OpenAI Gym convention.
[ 308 ]
Chapter 7
13. If the game is not over, the environment will return the next state, a reward,
and an indicator that the game is not over, along with an empty information
dictionary.
14. This function loads the daily returns of a universe of 100 random stocks.
15. The selector moves in random order over the files containing stock prices.
Some of them are corrupted so that loading them will result in an error.
The loader keeps trying until it has 100 pandas DataFrames containing stock
prices assembled. Only closing prices starting in 2005 will be considered.
16. In the next step, all DataFrames are concatenated. The percentile change
in stock price is calculated. All missing values are filled with zero, for no
change. Finally, we extract the values from the data as a NumPy array.
17. The last thing to do is transform the data into a time series. The first 100 steps
are the basis for the agent's decision. The 101st element is the next day's
return, on which the agent will be evaluated.
We only have to make minor edits in the A2CAgent agent class. Namely, we only
have to modify the model so that it can take in the time series of returns. To this
end, we add two LSTM layers, which actor and critic share:
def build_model(self):
state = Input(batch_shape=(None, #1
self.state_seq_length,
self.state_size))
x = LSTM(120,return_sequences=True)(state) #2
x = LSTM(100)(x)
mu = Dense(self.action_size, activation='tanh', #4
kernel_initializer='he_uniform')(actor_input)
[ 309 ]
Reinforcement Learning for Financial Markets
kernel_initializer='he_uniform')
(critic_input)
actor._make_predict_function()
critic._make_predict_function()
actor.summary()
critic.summary()
Again, we have built a Keras model in a function. It is only slightly different from
the model before. Let's explore it:
In the following diagram, you can see a visualization of how evolution strategies
work:
[ 310 ]
Chapter 7
Evolutionary strategy
To get a better grip on how this works, consider the following example. We want to
find a vector that minimizes the mean squared error to a solution vector. The learner
is not given the solution, but only the total error as a reward signal:
solution = np.array([0.5, 0.1, -0.3])
def f(w):
reward = -np.sum(np.square(solution - w))
return reward
[ 311 ]
Reinforcement Learning for Financial Markets
for j in range(npop): #4
w_try = w + N[j]
R[j] = f(w_try)
A = (R - np.mean(R)) / np.std(R) #5
w = w + alpha * np.dot(N.T, A)/npop #6
[ 312 ]
Chapter 7
{
Us ,a ( X ) = sup m ∈ R| E s ~ pπ ,a ~ π(s) u ( X − m) ≥ 0 }
Here U is a concave, continuous, and strictly increasing function that can be freely
chosen according to how much risk the trader is willing to take. The RL algorithm
now maximizes as follows:
J (π ) = U V ( s0 )
[ 313 ]
Reinforcement Learning for Financial Markets
The goal of the reward predictor is to produce a reward function that results in
a policy that has a large human preference. Human preference is measured by
showing the human the results of two policies and letting the human indicate
which one is more preferable:
[ 314 ]
Chapter 7
Robust RL
Much like for GANs, RL can be fragile and can be hard to train for good results. RL
algorithms are quite sensitive to hyperparameter choices. But there are a few ways
to make RL more robust:
• Using a larger experience replay buffer: The goal of using experience replay
buffers is to collect uncorrelated experiences. This can be achieved by just
creating a larger buffer or a whole buffer database that can store millions
of examples, possibly from different agents.
• Target networks: RL is unstable in part because the neural network
relies on its own output for training. By using a frozen target network
for generating training data, we can mitigate problems. The frozen target
network should only be updated slowly by, for example, moving the
weights of the target network only a few percent every few epochs in
the direction of the trained network.
• Noisy inputs: Adding noise to the state representation helps the model
generalize to other situations and avoids overfitting. It has proven especially
useful if the agent is trained in a simulation but needs to generalize to the
real, more complex world.
• Adversarial examples: In a GAN-like setup, an adversarial network can be
trained to fool the model by changing the state representations. The model
can, in turn, learn to ignore the adversarial attacks. This makes learning more
robust.
• Separating policy learning from feature extraction: The most well-known
results in reinforcement learning have learned a game from raw inputs.
However, this requires the neural network to interpret, for example, an
image by learning how that image leads to rewards. It is easier to separate
the steps by, for example, first training an autoencoder that compresses state
representations, then training a dynamics model that can predict the next
compressed state, and then training a relatively small policy network from
the two inputs.
Similar to the GAN tips, there is little theoretical reason for why these tricks work,
but they will make your RL work better in practice.
Frontiers of RL
You have now seen the theory behind and application of the most useful RL
techniques. Yet, RL is a moving field. This book cannot cover all of the current
trends that might be interesting to practitioners, but it can highlight some that
are particularly useful for practitioners in the financial industry.
[ 315 ]
Reinforcement Learning for Financial Markets
Multi-agent RL
Markets, by definition, include many agents. Lowe and others, 2017, Multi-Agent
Actor-Critic for Mixed Cooperative-Competitive Environments (see https://arxiv.org/
abs/1706.02275), shows that reinforcement learning can be used to train agents that
cooperate, compete, and communicate depending on the situation.
Multiple agents (in red) working together to chase the green dots. From the OpenAI blog.
In a task where the agent had to collect reward tokens, agents collaborated as
long as plenty of tokens were available and showed competitive behavior as
tokens got sparse. Zheng and others, 2017, MAgent: A Many-Agent Reinforcement
Learning Platform for Artificial Collective Intelligence (see https://arxiv.org/
abs/1712.00600), scaled the environment to include hundreds of agents. They
showed that agents developed more complex strategies such as an encirclement
attack on other agents through a combination of RL algorithms and clever reward
shaping.
[ 316 ]
Chapter 7
As you can see in the preceding diagram, the networks developed by NEAT are
often smaller than traditional, layer-based neural networks. They are hard to come
up with. This is the strength of AutoML; it can find effective strategies that humans
would not have discovered.
An alternative to using evolutionary algorithms for network design is to use
reinforcement learning, which yields similar results. There are a couple "off-the-
shelf" AutoML solutions:
• tpot (https://github.com/EpistasisLab/tpot): This is a data science
assistant that optimizes machine learning pipelines using genetic algorithms.
It is built on top of scikit-learn, so it does not create deep learning models but
models useful for structured data, such as random forests.
[ 317 ]
Reinforcement Learning for Financial Markets
For the subfield of hyperparameter search, there are a few packages available as well:
AutoML is still an active field of research, but it holds great promise. Many firms
struggle to use machine learning due to a lack of skilled employees. If machine
learning could optimize itself, more firms could start using machine learning.
Similarly, Banino and others in 2018 published a report titled, Vector-based navigation
using grid-like representations in artificial agents (see https://doi.org/10.1038/
s41586-018-0102-6), where they replicated so-called "grid cells" that allow
mammals to navigate using reinforcement learning.
The method is similar because both papers train RL algorithms on tasks related to the
area of research, for example, navigation. They then examine the learned weights of
the model for emergent properties. Such insight can be used to create more capable
RL agents but also to further the field of neuroscience.
[ 318 ]
Chapter 7
As the world of economics gets to grips with the idea that humans are not rational,
but irrational in predictable ways, understanding the brain becomes more important
when understanding economics. The results of neuroeconomics are particularly
relevant to finance as they deal with how humans act under uncertainty and
deal with risk, such as why humans are loss averse. Using RL is a promising
avenue to yield further insight into human behavior.
Exercises
As we've now completed the task, let's try our hand at two appropriate exercises
based on the content that we've covered.
Summary
In this chapter, you learned about the main algorithms in RL, Q-learning, policy
gradients, and evolutionary strategies. You saw how these algorithms could be
applied to trading and learned about some of the pitfalls of applying RL. You also
saw the direction of current research and how you can benefit from this research
today. At this point in the book, you are now equipped with a number of advanced
machine learning algorithms, which are hopefully useful to you when developing
machine learning models.
In the next chapter, we will discuss the practicalities of developing, debugging, and
deploying machine learning systems. We will break out of the data-science sandbox
and get our models into the real world.
[ 319 ]
Privacy, Debugging, and
Launching Your Products
Over the course of the last seven chapters we've developed a large toolbox of
machine learning algorithms that we could use for machine learning problems
in finance. To help round-off this toolbox, we're now going to look at what you
can do if your algorithms don't work.
Machine learning models fail in the worst way: silently. In traditional software,
a mistake usually leads to the program crashing, and while they're annoying for the
user, they are helpful for the programmer. At least it's clear that the code failed, and
often the developer will find an accompanying crash report that describes what went
wrong. Yet as you go beyond this book and start developing your own models, you'll
sometimes encounter machine learning code crashes too, which, for example, could
be caused if the data that you fed into the algorithm had the wrong format or shape.
These issues can usually be debugged by carefully tracking which shape the
data had at what point. More often, however, models that fail just output poor
predictions. They'll give no signal that they have failed, to the point that you might
not even be aware that they've even failed at all, but at other times, the model might
not train well, it won't converge, or it won't achieve a low loss rate.
In this chapter, we'll be focusing on how you debug these silent failures so that they
don't impact the machine learning algorithms that you've created. This will include
looking at the following subject areas:
• Finding flaws in your data that lead to flaws in your learned model
• Using creative tricks to make your model learn more from less data
• Unit testing data in production or training to ensure standards are met
[ 321 ]
Privacy, Debugging, and Launching Your Products
Debugging data
You'll remember that back in the first chapter of this book, we discussed how
machine learning models are a function of their training data, meaning that, for
example, bad data will lead to bad models, or as we put it, garbage in, garbage
out. If your project is failing, your data is the most likely culprit. Therefore, in this
chapter we will start by looking at the data first, before moving on to look at the
other possible issues that might cause our model to crash.
However, even if you have a working model, the real-world data coming in might not
be up to the task. In this section, we will learn how to find out whether you have good
data, what to do if you have not been given enough data, and how to test your data.
[ 322 ]
Chapter 8
To find out whether your model does contain predicting information, also called a
signal, you could ask yourself the question, could a human make a prediction given
this data? It's important for your AI to be given data that can be comprehended by
humans, because after all, the only reason we know intelligence is possible is because
we observe it in humans. Humans are good at understanding written text, but if a
human cannot understand a text, then the chances are that your model won't make
much sense of it either.
A common pitfall to this test is that humans have context that your model does
not have. A human trader does not only consume financial data, but they might have
also experienced the product of a company or seen the CEO on TV. This external
context flows into the trader's decision but is often forgotten when a model is built.
Likewise, humans are also good at focusing on important data. A human trader will
not consume all of the financial data out there because most of it is irrelevant.
Adding more inputs to your model won't make it better; on the contrary, it often
makes it worse, as the model overfits and gets distracted by all the noise. On the
other hand, humans are irrational; they follow peer pressure and have a hard time
making decisions in abstract and unfamiliar environments. Humans would struggle
to find an optimal traffic light policy, for instance, because the data that traffic lights
operate on is not intuitive to us.
This brings us to the second sanity check: a human might not be able to make
predictions, but there might be a causal (economic) rationale. There is a causal
link between a company's profits and its share price, the traffic on a road and traffic
jams, customer complaints and customers leaving your company, and so on. While
humans might not have an intuitive grasp of these links, we can discover them
through reasoning.
There are some tasks for which a causal link is required. For instance, for a long time,
many quantitative trading firms insisted on their data having a causal link to the
predicted outcomes of models. Yet nowadays, the industry seems to have slightly
moved away from that idea as it gets more confident in testing its algorithms. If
humans cannot make a prediction and there is no causal rationale for why your
data is predictive, you might want to reconsider whether your project is feasible.
Once you have determined that your data contains enough signal, you need to ask
yourself whether you have enough data to train a model to extract the signal. There
is no clear answer to the question of how much is enough, but roughly speaking, the
amount needed depends on the complexity of the model you hope to create. There
are a couple of rules of thumb to follow, however:
• For classification, you should have around 30 independent samples per class.
[ 323 ]
Privacy, Debugging, and Launching Your Products
• You should have 10 times as many samples as there are features, especially
for structured data problems.
• Your dataset should get bigger as the number of parameters in your model
gets bigger.
Keep in mind these rules are only rules of thumb and might be very different for
your specific application. If you can make use of transfer learning, then you can
drastically reduce the number of samples you need. This is why most computer
vision applications use transfer learning.
If you have any reasonable amount of data, say, a few hundred samples, then you
can start building your model. In this case, a sensible suggestion would be to start
with a simple model that you can deploy while you collect more data.
Most of the time, one of the best options would be to "augment your data."
We've already seen some data augmentation in Chapter 3, Utilizing Computer Vision.
Of course, you can augment all kinds of data in various ways, including slightly
changing some database entries. Taking augmentation a step further, you might be
able to generate your data, for example, in a simulation. This is effectively how most
reinforcement learning researchers gather data, but this can also work in other cases.
The data we used for fraud detection back in Chapter 2, Applying Machine Learning to
Structured Data was obtained from simulation. The simulation requires you to be able
to write down the rules of your environment within a program. Powerful learning
algorithms tend to figure out these often over-simplistic rules, so they might not
generalize to the real world as well. Yet, simulated data can be a powerful addition
to real data.
Likewise, you can often find external data. Just because you haven't tracked a certain
data point, it does not mean that nobody else has. There is an astonishing amount of
data available on the internet. Even if the data was not originally collected for your
purpose, you might be able to retool data by either relabeling it or by using it for
transfer learning. You might be able to train a model on a large dataset for a different
task and then use that model as a basis for your task. Equally, you can find a model
that someone else has trained for a different task and repurpose it for your task.
[ 324 ]
Chapter 8
Finally, you might be able to create a simple model, which does not capture the
relationship in the data completely but is enough to ship a product. Random forests
and other tree-based methods often require much less data than neural networks.
It's important to remember that for data, quality trumps quantity in the majority
of cases. Getting a small, high-quality dataset in and training a weak model is
often your best shot to find problems with data early. You can always scale up data
collection later. A mistake many practitioners make is that they spend huge amounts
of time and money on getting a big dataset, only to find that they have the wrong
kind of data for their project.
Additionally, you need to test whether your data is clean from things such as
personal information. As we'll see in the following section on privacy, personal
information is a liability that you want to get rid of, unless you have good reasons
and consent from the user to use it.
Since monitoring data quality is important when trading based on many data
sources, Two Sigma Investments LP, a New York City-based international hedge
fund, has created an open source library for data monitoring. It is called marbles, and
you can read more about it here: https://github.com/twosigma/marbles. marbles
builds on Python's unittest library.
The following code sample shows a simple marbles unit test. Imagine you are
gathering data about the unemployment rate in Ireland. For your models to work,
you need to ensure that you actually get the data for consecutive months, and don't
count one month twice, for instance.
[ 325 ]
Privacy, Debugging, and Launching Your Products
import pandas as pd #2
import numpy as np
from datetime import datetime, timedelta
class TimeSeriesTestCase(marbles.core.TestCase,
mixins.MonotonicMixins): #3
def setUp(self): #4
self.df = pd.DataFrame({'dates':[datetime(2018,1,1),
datetime(2018,2,1),
datetime(2018,2,1)],
'ireland_unemployment':[6.2,6.1,6.0]}) #5
def tearDown(self):
self.df = None #6
def test_date_order(self): #7
self.assertMonotonicIncreasing(sequence=self.df.dates,
note = 'Dates need to increase
monotonically') #8
Don't worry if you don't fully understand the code. We're now going to go through
each stage of the code:
1. Marbles features two main components. The core module does the actual
testing, while the mixins module provides a number of useful tests for
different types of data. This simplifies your test writing and gives you more
readable and semantically interpretable tests.
2. You can use all the libraries, like pandas, that you would usually use to
handle and process data for testing.
3. Now it is time to define our test class. A new test class must inherit marbles'
TestCase class. This way, our test class is automatically set up to run as
a marbles test. If you want to use a mixin, you also need to inherit the
corresponding mixin class.
In this example, we are working with a series of dates that should be increasing
monotonically. The MonotonicMixins class provides a range of tools that
allow you to test for a monotonically increasing series automatically.
[ 326 ]
Chapter 8
To run a unit test in a Jupyter Notebook, we need to tell marbles to ignore the first
argument; we achieve this by running the following:
if __name__ == '__main__':
marbles.core.main(argv=['first-arg-is-ignored'], exit=False)
It's more common to run unit tests directly from the command line. So, if you saved
the preceding code in the command line, you could run it with this command:
python -m marbles marbles_test.py
Of course, there are problems with our data. Luckily for us, our test ensures that this
error does not get passed on to our model, where it would cause a silent failure in the
form of a bad prediction. Instead, the test will fail with the following error output:
[ 327 ]
Privacy, Debugging, and Launching Your Products
F #1
==================================================================
FAIL: test_date_order (__main__.TimeSeriesTestCase) #2
------------------------------------------------------------------
marbles.core.marbles.ContextualAssertionError: Elements in 0
2018-01-01
1 2018-02-01
2 2018-02-01 #3
Name: dates, dtype: datetime64[ns] are not strictly monotonically
increasing
Source (<ipython-input-1-ebdbd8f0d69f>): #4
19
> 20 self.assertMonotonicIncreasing(sequence=self.df.dates,
21 note = 'Dates need to increase
monotonically')
22
Locals: #5
Note: #6
Dates need to increase monotonically
----------------------------------------------------------------------
FAILED (failures=1)
So, what exactly caused the data to fail? Let's have a look:
1. The top line shows the status of the entire test. In this case, there was only one
test method, and it failed. Your test might have multiple different test methods,
and marbles would display the progress by showing how tests fail or pass.
The
next couple of lines describe the failed test method. This line describes that
the test_date_order method of the TimeSeriesTestCase class failed.
2. marbles shows precisely how the test failed. The values of the dates tested
are shown, together with the cause for failure.
3. In addition to the actual failure, marbles will display a traceback showing
the actual code where our test failed.
4. A special feature of marbles is the ability to display local variables. This way,
we can ensure that there was no problem with the setup of the test. It also
helps us in getting the context as to how exactly the test failed.
[ 328 ]
Chapter 8
5. Finally, marbles will display our note, which helps the test consumer
understand what went wrong.
6. As a summary, marbles displays that the test failed with one failure.
Sometimes, you may be able to accept data even though it failed some tests,
but more often than not you'll want to dig in and see what is going on.
The point of unit testing data is to make the failures loud in order to prevent data
issues from giving you bad predictions. A failure with an error message is much
better than a failure without one. Often, the failure is caused by your data vendor,
and by testing all of the data that you got from all of the vendors, it will allow you
to be aware when a vendor makes a mistake.
Unit testing data also helps you to ensure you have no data that you shouldn't have,
such as personal data. Vendors need to clean datasets of all personally identifying
information, such as social security numbers, but of course, they sometimes forget.
Complying with ever stricter data privacy regulation is a big concern for many
financial institutions engaging in machine learning.
The next section will therefore discuss how to preserve privacy and comply with
regulations while still gaining benefits from machine learning.
This text will not go into depth on how to comply with this law specifically.
However, if you wish to expand your understanding of the topic, then the UK
government's guide to GDPR is a good starting place to learn more about the
specifics of the regulation and how to comply with it: https://www.gov.uk/
government/publications/guide-to-the-general-data-protection-
regulation.
This section will outline both the key principles of the recent privacy legislation
and some technological solutions that you can utilize in order to comply with these
principles.
[ 329 ]
Privacy, Debugging, and Launching Your Products
The overarching rule here is to, "delete what you don't need." For a long time, a large
percentage of companies have just stored all of the data that they could get their hands
on, but this is a bad idea. Storing personal data is a liability for your business. It's
owned by someone else, and you are on the hook for taking care of it. The next time
you hear a statement such as, "We have 500,000 records in our database," think of it
more along the lines of, "We have 500,000 liabilities on our books." It can be a good idea
to take on liabilities, but only if there is an economic value that justifies these liabilities.
What happens astonishingly often though is that you might collect personal data by
accident. Say you are tracking device usage, but accidentally include the customer ID
in your records. You need practices in place that monitor and prevent such accidents,
here are four of the key ones:
[ 330 ]
Chapter 8
• Encrypt and Obfuscate data: Apple, for instance, collects phone data but
adds random noise to the collected data. The noise renders each individual
record incorrect, but in aggregate the records still give a picture of user
behavior. There are a few caveats to this approach; for example, you can only
collect so many data points from a user before the noise cancels out, and the
individual behavior is revealed.
Noise, as introduced by obfuscation, is random. When averaged over a
large sample of data about a single user, the mean of the noise will be zero
as it does not present a pattern by itself. The true profile of the user will be
revealed. Similarly, recent research has shown that deep learning models can
learn on homomorphically encrypted data. Homomorphic encryption is a
method of encryption that preserves the underlying algebraic properties of
the data. Mathematically, this can be expressed as follows:
D ( E (m1 + m2 )) = m1 + m2
• Train locally, and upload only a few gradients: One way to avoid uploading
user data is to train your model on the user's device. The user accumulates
data on the device. You can then download your model on to the device
and perform a single forward and backward pass on the device.
To avoid the possibility of inference of user data from the gradients, you only
upload a few gradients at random. You can then apply the gradients to your
master model.
[ 331 ]
Privacy, Debugging, and Launching Your Products
To further increase the overall privacy of the system, you do not need to
download all the newly update weights from the master model to the user's
device, but only a few. This way, you train your model asynchronously
without ever accessing any data. If your database gets breached, no user
data is lost. However, we need to note that this only works if you have
a large enough user base.
• Standardization ensures that all of the data has a mean of zero and
a standard deviation of one. It is computed by subtracting the mean
and dividing by the standard deviation of the data:
x−µ
x′ =
σ
This is probably the most common way of scaling features. It's especially
useful if you suspect that your data contains outliers as it is quite robust. On
the flip side, standardization does not ensure that your features are between
zero and one, which is the range in which neural networks learn best.
• Min-Max rescaling does exactly that. It scales all data between zero and
one by first subtracting the minimum value and then dividing by the range
of values. We can see this expressed in the formula below:
x − min ( x)
x′ =
max ( x) − min ( x)
If you know for sure that your data contains no outliers, which is the case in
images, for instance, Min-Max scaling will give you a nice scaling of values
between zero and one.
[ 332 ]
Chapter 8
x−µ
x′ =
max ( x) − min ( x)
• For some applications, it is better to not scale individual features, but instead
vectors of features. In this case, you would apply unit length scaling by
dividing each element in the vector by the total length of the vector, as we
can see below:
x
x′ =
x
The length of the vector usually means the L2 norm of the vector x 2 , that is,
the square root of the sum of squares. For some applications, the vector length
means the L1 norm of the vector, x 1 , which is the sum of vector elements.
However you scale, it is important to only measure the scaling factors, mean, and
standard deviation on the test set. These factors include only a select amount of the
information about the data. If you measure them over your entire dataset, then the
algorithm might perform better on the test set than it will in production, due to this
information advantage.
Equally importantly, you should check that your production code has proper feature
scaling as well. Over time, you should recalculate your feature distribution and
adjust your scaling.
[ 333 ]
Privacy, Debugging, and Launching Your Products
Note: The authors of LIME, Ribeiro, Singh, and Guestrin, curated a great
GitHub repository around their algorithm with many explanations and
tutorials, which you can find here: https://github.com/marcotcr/lime.
On Kaggle kernels, LIME is installed by default. However, you can install LIME
locally with the following command:
pip install lime
The LIME algorithm works with any classifier, which is why it is model agnostic.
To make an explanation, LIME cuts up the data into several sections, such as areas
of an image or utterances in a text. It then creates a new dataset by removing some
of these features. It runs this new dataset through the black box classifier and obtains
the classifiers predicted probabilities for different classes. LIME then encodes the
data as vectors describing what features were present. Finally, it trains a linear model
to predict the outcomes of the black box model with different features removed. As
linear models are easy to interpret, LIME will use the linear model to determine the
most important features.
Let's say that you are using a text classifier, such as TF-IDF, to classify emails such
as those in the 20 newsgroup dataset. To get explanations from this classifier, you
would use the following snippet:
from lime.lime_text import LimeTextExplainer #1
explainer = LimeTextExplainer(class_names=class_names) #2
exp = explainer.explain_instance(test_example, #3
classifier.predict_proba, #4
num_features=6) #5
exp.show_in_notebook() #6
1. The LIME package has several classes for different types of data.
2. To create a new blank explainer, we need to pass the names of classes
of our classifier.
3. We'll provide one text example for which we want an explanation.
4. We provide the prediction function of our classifier. We need to provide
a function that provides probabilities. For Keras, this is just model.predict;
for scikit models, we need to use the predict_proba method.
5. LIME shows the maximum number of features. We want to show only
the importance of the six most important features in this case.
6. Finally, we can render a visualization of our prediction, which looks like this:
[ 334 ]
Chapter 8
The explanation shows the classes with different features that the text gets classified
as most often. It shows the words that most contribute to the classification in the
two most frequent classes. Under that, you can see the words that contributed to
the classification highlighted in the text.
As you can see, our model picked up on parts of the email address of the sender as
distinguishing features, as well as the name of the university, "Rice." It sees "Caused"
to be a strong indicator that the text is about atheism. Combined, these are all things
we want to know when debugging datasets.
LIME does not perfectly solve the problem of explaining models. It struggles if the
interaction of multiple features leads to a certain outcome for instance. However,
it does well enough to be a useful data debugging tool. Often, models pick up on
things they should not be picking up on. To debug a dataset, we need to remove
all these "give-away" features that statistical models like to overfit to.
Looking back at this section, you've now seen a wide range of tools that you can use
to debug your dataset. Yet, even with a perfect dataset, there can be issues when it
comes to training. The next section is about how to debug your model.
[ 335 ]
Privacy, Debugging, and Launching Your Products
There are a number of different hyperparameter optimization tools, but we will look
at Hyperas because of its ease of use. Hyperas is a wrapper for hyperopt, a popular
optimization library made for working with Keras.
Depending on your setup, you might need to make a few adjustments to the
installation. If this is the case, then the Hyperas GitHub page, link above, offers
more information.
Hyperas offers two optimization methods, Random Search and Tree of Parzen
Estimators. Within a range of parameters that we think are reasonable, the random
search will sample randomly and train a model with random hyperparameters.
It will then pick the best-performing model as the solution.
Random search is simple and robust, and it can be scaled easily. It basically makes
no assumption about the hyperparameters, their relation, and the loss surface. On
the flip side, it is relatively slow.
The Tree of Parzen (TPE) algorithm models the relation P(x|y), where x represents
the hyperparameters and y the associated performance. This is the exact opposite
modeling of Gaussian processes, which model P(y|x) and are popular with many
researchers.
[ 336 ]
Chapter 8
Empirically, it turns out that TPE performs better. For the precise details, see the 2011
paper, Algorithms for Hyper-Parameter Optimization, available at: https://papers.
nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization -- that
was authored by James S. Bergstra and others. TPE is faster than random search
but can get stuck in local minima and struggles with some difficult loss surfaces.
As a rule of thumb, it makes sense to start with TPE, and if TPE struggles, move
to random search.
Note: The code for this example can be found at: https://www.
kaggle.com/jannesklaas/Hyperas.
The following example will show you how to use Hyperas and Hyperopt for an
MNIST dataset classifier:
from hyperopt import Trials, STATUS_OK, tpe #1
from hyperas import optim #2
from hyperas.distributions import choice, uniform
While the code was short, let's explain what it all means:
To build on the previous ideas that we've explored, let's now add the following,
which we will explain in more detail once the code has been written:
def data(): #1
import numpy as np #2
from keras.utils import np_utils
path = '../input/mnist.npz' #3
with np.load(path) as f:
X_train, y_train = f['x_train'], f['y_train']
[ 337 ]
Privacy, Debugging, and Launching Your Products
1. Hyperas expects a function that loads the data; we cannot just pass on
a dataset from memory.
2. To scale the search, Hyperas creates a new runtime in which it does model
creation and evaluation. This also means imports that we did in a notebook
do not always transfer into the runtime. To be sure that all modules are
available, we need to do all imports in the data function. This is also true
for modules that will only be used for the model.
3. We now load the data. Since Kaggle kernels do not have access to the
internet, we need to load the MNIST data from disk. If you have internet,
but no local version of the files, you can get the data using following code:
from keras.datasets import mnist
(Y_train, y_train), (X_test, y_test) = mnist.load_data()
I would still keep the no internet version around because it is the default
setting.
4. The data function also needs to preprocess the data. We do the standard
reshaping and scaling that we did when we worked with MNIST earlier.
5. Finally, we return the data. This data will be passed into the function that
builds and evaluates the model:
def model(X_train, y_train, X_test, y_test): #1
model = Sequential() #2
model.add(Dense(512, input_shape=(784,)))
model.add(Activation('relu'))
model.add(Dropout({{uniform(0, 0.5)}})) #3
[ 338 ]
Chapter 8
model.add(Activation({{choice(['relu','tanh'])}})) #5
model.add(Dropout({{uniform(0, 0.5)}}))
model.add(Dense(10))
model.add(Activation('softmax'))
rms = RMSprop()
model.compile(loss='categorical_crossentropy',
optimizer=rms,
metrics=['accuracy'])
model.fit(X_train, y_train, #6
batch_size={{choice([64, 128])}},
epochs=1,
verbose=2,
validation_data=(X_test, y_test))
score, acc = model.evaluate(X_test, y_test, verbose=0) #7
print('Test accuracy:', acc)
return {'loss': -acc, 'status': STATUS_OK, 'model': model} #8
As you can see, the preceding snippet of code is made up of eight defining pieces.
Let's now explore them so that we're able to fully understand the code we've just
produced:
1. The model function both defines the model and evaluates it. Given a training
dataset from the data function, it returns a set of quality metrics.
2. When fine-tuning with Hyperas, we can define a Keras model just as we
usually would. Here, we only have to replace the hyperparameters we want
to tune with Hyperas functions.
3. To tune dropout, for instance, we replace the Dropout hyperparameter with
{{uniform(0, 0.5)}}. Hyperas will automatically sample and evaluate
dropout rates between 0 and 0.5, sampled from a uniform distribution.
4. To sample from discrete distributions, for instance, the size of a hidden layer,
we replace the hyperparameter with {{choice([256, 512, 1024])}}.
Hyperas will choose from a hidden layer size of 256, 512, and 1,024 now.
5. We can do the same to choose activation functions.
6. To evaluate the model, we need to compile and fit it. In this process, we can
also choose between different batch sizes. In this case, we only train for one
epoch, to keep the time needed for this example short. You could also run
a whole training process with Hyperas.
7. To gain insight into how well the model is doing, we evaluate it on test data.
[ 339 ]
Privacy, Debugging, and Launching Your Products
8. Finally, we return the model's score, the model itself, and an indicator
that everything went okay. Hyperas tries to minimize a loss function. To
maximize accuracy, we set the loss to be the negative accuracy. You could
also pass the model loss here, depending on what the best optimization
method is for your problem.
We pass the model method and the data method, and we specify how many trials
we want to run and which class should govern the trials. Hyperopt also offers
a distributed trials class in which workers communicate via MongoDB.
When working in a Jupyter Notebook, we need to provide the name of the notebook
we are working in. Kaggle Notebooks all have the filename __notebook_source__,
regardless of the name you gave them.
After it's run, Hyperas returns the best-performing model as well as the
hyperparameters of the best model. If you print out best_run, you should
see output similar to this:
{'Activation': 1,
'Dense': 1,
'Dropout': 0.3462695171578595,
'Dropout_1': 0.10640021656377913,
'batch_size': 0}
For choice selections, Hyperas shows the index. In this case, the activation function
tanh was chosen.
In this case we ran the hyperparameter search only for a few trials. Usually, you
would run a few hundred or thousand trials. To do this we would use automated
hyperparameter search, which can be a great tool to improve model performance
if you have enough compute power available.
However, it won't get a model that does not work at all to work. When choosing this
approach, you need to be sure to have a somewhat-working approach first before
investing in hyperparameter search.
[ 340 ]
Chapter 8
When doing a line search, it is better to increase the learning rate exponentially. You
are more likely to care about the region of smaller learning rates than about very
large learning rates.
In our example below, we perform 20 evaluations and double the learning rate
in every evaluation. We can run this by executing the following code:
init_lr = 1e-6 #1
losses = []
lrs = []
for i in range(20): #2
model = Sequential()
model.add(Dense(512, input_shape=(784,)))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.2))
model.add(Dense(10))
model.add(Activation('softmax'))
opt = Adam(lr=init_lr*2**i) #3
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['acc'])
loss = hist.history['loss'][0] #5
losses.append(loss)
lrs.append(init_lr*2**i)
[ 341 ]
Privacy, Debugging, and Launching Your Products
Let's now take a more detailed look at the preceding featured code:
1. We specify a low, but still reasonable, initial learning rate from which we
start our search.
2. We then perform training 20 times with different learning rates. We need
to set up the model from scratch each time.
3. We calculate our new learning rate. In our case, we double the learning rate
in each evaluation step. You could also use a smaller increase if you want
a more fine-grained picture.
4. We then fit the model with our new learning rate.
5. Finally, we keep track of the loss.
If your dataset is very large, you can perform this learning rate search on a subset
of the data. The interesting part comes from the visualization of learning rates:
fig, ax = plt.subplots(figsize = (10,7))
plt.plot(lrs,losses)
ax.set_xscale('log')
When you run this code, it will then output the following chart:
[ 342 ]
Chapter 8
As you can see, the loss is optimal between 1e-3 and 1e-2. We can also see that
the loss surface is relatively flat in this area. This gives us insight that we should
use a learning rate around 1e-3. To avoid overshooting, we select a learning rate
somewhat lower than the optimum found by line search.
Here l is the cycle length and a0 is the initial learning rate. We modify this function
to ensure that t does not become larger than the cycle length:
def cosine_anneal_schedule(t):
lr_init = 1e-2 #1
anneal_len = 5
if t >= anneal_len: t = anneal_len -1 #2
cos_inner = np.pi * (t % (anneal_len)) #3
cos_inner /= anneal_len
cos_out = np.cos(cos_inner) + 1
return float(lr_init / 2 * cos_out)
[ 343 ]
Privacy, Debugging, and Launching Your Products
To get a better understanding of what the learning rate scheduling function does,
we can plot the learning rate it would set over 10 epochs:
srs = [cosine_anneal_schedule(t) for t in range(10)]
plt.plot(srs)
With the output of the code being shown in the following graph:
Cosine anneal
We now have a callback that Keras will call at the end of each epoch in order to get
a new learning rate. We pass this callback to the fit method and voilà, our model
trains with a decreasing learning rate:
model.fit(x_train,y_train,batch_size=128,epochs=5,callbacks=[cb])
A version of the learning rate annealing is to add restarts. At the end of an annealing
cycle, we move the learning rate back up. This is a method used to avoid overfitting.
With a small learning rate, our model might find a very narrow minimum. If the data
we want to use our model on is slightly different from the training data, then the
loss surface might change a bit, and our model could be out of the narrow minimum
for this new loss surface. If we set the learning rate back up, our model will get out
of narrow minima. Broad minima, however, are stable enough for the model to stay
in them:
[ 344 ]
Chapter 8
As the cosine function goes back up by itself, we only have to remove the line to stop
it from doing so:
def cosine_anneal_schedule(t):
lr_init = 1e-2
anneal_len = 10
cos_inner = np.pi * (t % (anneal_len))
cos_inner /= anneal_len
cos_out = np.cos(cos_inner) + 1
return float(lr_init / 2 * cos_out)
[ 345 ]
Privacy, Debugging, and Launching Your Products
To provide an interface from which you can watch your model's progress,
TensorBoard also offers some options useful for debugging. For example, you
can observe the distributions of the model's weights and gradients during training.
To use TensorBoard with Keras, we set up a new callback. TensorBoard has many
options, so let's walk through them step by step:
from keras.callbacks import TensorBoard
tb = TensorBoard(log_dir='./logs/test2', #1
histogram_freq=1, #2
batch_size=32, #3
write_graph=True, #4
write_grads=True,
write_images=True,
embeddings_freq=0, #5
embeddings_layer_names=None,
embeddings_metadata=None)
There are five key pieces of the preceding code that we need to take into
consideration:
1. First, we need to specify where Keras should save the data that TensorBoard
later visualizes. Generally, it is a good idea to save all logs of your different
runs in one logs folder and give every run its own subfolder, such as
test2 in this case. This way, you can easily compare different runs within
TensorBoard but also keep different runs separate.
2. By default, TensorBoard would just show you the loss and accuracy
of your model. In this case, we are interested in histograms showing
weights and distributions. We save the data for the histograms every epoch.
3. To generate data, TensorBoard runs batches through the model. We need
to specify a batch size for this process.
[ 346 ]
Chapter 8
Once we have the callback set up, we can pass it to the training process. We
will train the MNIST model once again. We multiply the inputs by 255, making
training much harder. To achieve all of this we need to run the following code:
hist = model.fit(x_train*255,y_train,
batch_size=128,
epochs=5,
callbacks=[tb],
validation_data=(x_test*255,y_test))
Here full_path_to_your_logs is the path you saved your logs in, for example,
logs in our case. TensorBoard runs on port 6006 by default, so in your browser,
go to http://localhost:6006 to see TensorBoard.
Once the page has loaded, navigate to the HISTOGRAMS section; this section
should look something like this:
TensorBoard histograms
[ 347 ]
Privacy, Debugging, and Launching Your Products
You can see the distribution of gradients and weights in the first layer. As you can
see, the gradients are uniformly distributed and extremely close to zero. The weights
hardly change at all over the different epochs. We are dealing with a vanishing
gradient problem; we will cover this problem in depth later.
Armed with the real-time insight that this problem is happening, we can react faster.
If you really want to dig into your model, TensorBoard also offers a visual debugger.
In this debugger, you can step through the execution of your TensorFlow model
and examine every single value inside it. This is especially useful if you are working
on complex models, such as generative adversarial networks, and are trying to
understand why something complex goes wrong.
Note: The TensorFlow debugger does not work well with models trained
in Jupyter Notebooks. Save your model training code to a Python .py
script and run that script.
To use the TensorFlow debugger, you have to set your model's runtime to a special
debugger runtime. In specifying the debugger runtime, you also need to specify
which port you want the debugger to run, in this case, port 2018:
import tensorflow as tf
from tensorflow.python import debug as tf_debug
import keras
keras.backend.set_session(
tf_debug.TensorBoardDebugWrapperSession(
tf.Session(), "localhost:2018"))
Once Keras begins to work with the debugger runtime, you can debug your model.
For the debugger to work, you need to name your Keras model to model. However,
you do not need to train the model with a TensorBoard callback.
Now, start TensorBoard and activate the debugger by specifying the debugger port
as follows:
tensorboard --logdir=/full_path_to_your_logs --debugger_port 2018
Now you can open TensorBoard as usual in your browser on port 6006. TensorBoard
now has a new section called DEBUGGER:
[ 348 ]
Chapter 8
TensorBoard debugger
By clicking STEP, you execute the next step in the training process. With
CONTINUE…, you can train your model for one or more epochs. By navigating
the tree on the left side, you can view the components of your model. You can
visualize individual elements of your model, to see how different actions affect
them. Using the debugger effectively requires a bit of practice, but if you are
working with complex models, it is a great tool.
Of the two, the vanishing gradient problem is the more persistent issue. Vanishing
gradients are caused by the fact that in deep networks, gradients of earlier layers
depend on gradients of layers closer to the output. If the output gradients are small,
then the gradients behind them are even smaller. Thus, the deeper the network, the
more issues that occur with regard to vanishing gradients.
The key causes of small gradients include sigmoid and tanh activation functions.
If you look at the following sigmoid function, you'll see that it is very flat toward
large values:
[ 349 ]
Privacy, Debugging, and Launching Your Products
Sigmoid vanishing
The small gradients of the sigmoid function are the reason why the ReLU activation
function has become popular for training deep neural networks. Its gradient is equal
to one for all positive input values. However, it is zero for all negative input values.
Another cause of vanishing gradients is saddle points in the loss function. Although
no minimum was reached, the loss function is very flat in some areas, producing
small gradients.
To combat the vanishing gradient problem, you should use ReLU activation. If you
see that your model is training slowly, consider increasing the learning rate to move
out of a saddle point faster. Finally, you might just want to let the model train longer
if it suffers from small gradients.
The exploding gradient problem is usually caused by large absolute weight values.
As backpropagation multiplies the later layers' gradients with the layers' weights,
large weights amplify gradients. To counteract the exploding gradient problem,
you can use weight regularization, which incentivizes smaller weights. Using
a method called gradient clipping, you can ensure that gradients do not become
larger than a certain value. In Keras, you can clip both the norm and the absolute
value of gradients:
[ 350 ]
Chapter 8
If exploding gradients are a problem, you can add a batch normalization layer
to your model as follows:
from keras.layers import BatchNormalization
model.add(BatchNormalization())
Batch normalization also reduces the risk of vanishing gradients and has enabled
the construction of much deeper networks recently.
You have now seen a wide range of tools that can be used to debug your models. As a
final step, we are going to learn some methods to run models in production and speed
up the machine learning process.
Deployment
Deployment into production is often seen as separate from the creation of models. At
many companies, data scientists create models in isolated development environments
on training, validation, and testing data that was collected to create models.
Once the model performs well on the test set, it then gets passed on to deployment
engineers, who know little about how and why the model works the way it does.
This is a mistake. After all, you are developing models to use them, not for the fun
of developing them.
Models tend to perform worse over time for several reasons. The world changes, so
the data you trained on might no longer represent the real world. Your model might
rely on the outputs of some other systems that are subject to change. There might be
unintended side effects and weaknesses of your model that only show with extended
usage. Your model might influence the world that it tries to model. Model decay
describes how models have a lifespan after which performance deteriorates.
Data scientists should have the full life cycle of their models in mind. They need
to be aware of how their model works in production in the long run.
[ 351 ]
Privacy, Debugging, and Launching Your Products
This section describes some best practices for getting your models to work in the
real world. The exact method of serving your model can vary depending on your
application. See the upcoming section Performance Tips for more details on choosing
a deployment method.
Launching fast
The process of developing models depends on real-world data as well as an insight
into how the performance of the model influences business outcomes. The earlier
you can gather data and observe how model behavior influences outcomes, the
better. Do not hesitate to launch your product with a simple heuristic.
Take the case of fraud detection, for instance. Not only do you need to gather
transaction data together with information about occurring frauds, you also want
to know how quick fraudsters are at finding ways around your detection methods.
You want to know how customers whose transactions have been falsely flagged as
fraud react. All of this information influences your model design and your model
evaluation metrics. If you can come up with a simple heuristic, deploy the heuristic
and then work on the machine learning approach.
When developing a machine learning model, try simple models first. A surprising
number of tasks can be modeled with simple, linear models. Not only do you obtain
results faster, but you can also quickly identify the features that your model is likely
to overfit to. Debugging your dataset before working on a complex model can save
you many headaches.
A second advantage of getting a simple approach out of the door quickly is that
you can prepare your infrastructure. Your infrastructure team is likely made up of
different people from the modeling team. If the infrastructure team does not have to
wait for the modeling team but can start optimizing the infrastructure immediately,
then you gain a time advantage.
You might predict whether the user is interested in a given product, measured by
the user reading the product description. However, the metric you want to optimize
in your application is not your model accuracy, but the click-through rate of users
going to the description screen. On a higher order, your business is not designed
to maximize the click-through rate, but revenue. If your users only click on
low-revenue products, your click-through rate does not help you.
Often, you will find that to align with a higher order metric, you'll need to combine
several metrics, such as accuracy and speed of predictions. In this case, you should
craft a formula that combines the metrics into one single number. A single number
will allow you to doubtlessly choose between two models and help your engineers
to create better models.
[ 353 ]
Privacy, Debugging, and Launching Your Products
For instance, you could set a maximum latency of 200 milliseconds and your metric
would be, "Accuracy if latency is below 200 milliseconds, otherwise zero." If you
do not wish to set one maximum latency value, you could choose, "Accuracy divided
by latency in milliseconds." The exact design of this formula depends on your
application. As you observe how your model influences its higher order metric,
you can adapt your model metric. The metric should be simple and easy to quantify.
Next, to regularly test your model's impact on higher order metrics, you
should regularly test your models own metrics, such as accuracy. To this end, you
need a constant stream of ground truth labels together with your data. In some cases,
such as detecting fraud, ground truth data is easily collected, although it might come
in with some latency. In this case, customers might need a few weeks to find out they
have been overcharged.
In other cases, you might not have ground truth labels. Often, you can hand-label
data for which you have no ground truth labels coming in. Through good UI design,
the process of checking model predictions can be fast. Testers only have to decide
whether your model's prediction was correct or not, something they can do through
button presses in a web or mobile app. If you have a good review system in place,
data scientists who work on the model should regularly check the model's outputs.
This way, patterns in failures (our model does poorly on dark images) can be
detected quickly, and the model can be improved.
If, say, the collection method of the data changes, the distribution of your data might
change too. This could break your model. Equally, the real world might just change,
and with it the data distribution. To avoid changes in the data breaking your model,
you first need to be aware of what data you are using and assign an owner to each
feature. The job of the feature owner is to investigate where the data is coming from
and alert the team if changes in the data are coming. The feature owner should
also write down the assumptions underlying the data. In the best case, you test
these assumptions for all new data streaming in. If the data does not pass the tests,
investigate and eventually modify your model.
Equally, your model outputs might get used as inputs of other models. Help
consumers of your data reach you by clearly identifying yourself as the owner
of the model.
[ 354 ]
Chapter 8
Alert users of your model of changes to your model. Before deploying a model,
compare the new model's predictions to the old model's predictions. Treat models
as software and try to identify "breaking changes," that would significantly alter
your model's behavior. Often, you might not know who is accessing your model's
predictions. Try to avoid this by clear communication and setting access controls
if necessary.
Just as software has dependencies, libraries that need to be installed for the software
to work, machine learning models have data dependencies. Data dependencies are
not as well understood as software dependencies. By investigating your model's
dependencies, you can reduce the risk of your model breaking when data changes.
Performance tips
In many financial applications, speed is of the essence. Machine learning, especially
deep learning, has a reputation for being slow. However, recently, there have been
many advances in hardware and software that enable faster machine learning
applications.
In natural language processing (NLP), for instance, batch sizes often need to be small,
so the parallelization of operations does not work as well since not that many samples
are processed at the same time. Additionally, some words appear much more often
than others, giving large benefits to caching frequent words. Thus, many NLP tasks
run faster on CPUs than GPUs. If you can work with large batches, however, a GPU
or even specialized hardware is preferable.
[ 355 ]
Privacy, Debugging, and Launching Your Products
From version 1.8 onward, the estimator API's features distribute training on multiple
machines, while the Keras API does not feature them yet. Estimators also have a
number of other speed-up tricks, so they are usually faster than Keras models.
You can find information on how to set up your cluster for distributed
TensorFlow here: https://www.tensorflow.org/deploy/
distributed.
By changing the import statements, you can easily use Keras as part of TensorFlow
and don't have to change your main code:
import tensorflow as tf
from tensorflow.python import keras
In this section, we will create a model to learn the MNIST problem before training
it using the estimator API. First, we load and prepare the dataset as usual. For more
efficient dataset loading, see the next section:
(x_train, y_train), (x_test, y_test) =
keras.datasets.mnist.load_data()
x_train.shape = (60000, 28 * 28)
x_train = x_train / 255
y_train = keras.utils.to_categorical(y_train)
model.compile(optimizer=keras.optimizers.SGD(lr=0.0001,
momentum=0.9),
loss='categorical_crossentropy',
metric='accuracy')
[ 356 ]
Chapter 8
To set up training, we need to know the name assigned to the model input. We can
quickly check this with the following code:
model.input_names
['dense_1_input']
Estimators get trained with an input function. The input function allows us to
specify a whole pipeline, which will be executed efficiently. In this case, we only
want an input function that yields our training set:
train_input_fn = tf.estimator.inputs.numpy_input_fn(
x={'dense_1_input': x_train},
y=y_train,
num_epochs=1,
shuffle=False)
Finally, we train the estimator on the input. And that is it; you can now utilize
distributed TensorFlow with estimators:
estimator.train(input_fn=train_input_fn, steps=2000)
When you lock in your model to specialized hardware, you can often make
significant gains in your performance. If you have the resources, it might even make
sense to write your own specialized layer in CUDA. If you want to change hardware
later, you can usually export weights and import them to a different layer.
[ 357 ]
Privacy, Debugging, and Launching Your Products
If your GPU utilization is not at around 80% to 100%, you can gain significantly
by optimizing your pipeline. There are several steps you can take to optimize
your pipeline:
You need to make sure your generator is thread safe. You can make any
generator thread safe with the following code snippet:
import threading
class thread_safe_iter: #1
def __init__(self, it):
self.it = it
self.lock = threading.Lock()
def __iter__(self):
return self
def next(self): #2
with self.lock:
return self.it.next()
def thread_safe_generator(f): #3
def g(*a, **kw):
return thread_safe_iter(f(*a, **kw))
return g
@thread_safe_generator
def gen():
[ 358 ]
Chapter 8
You can also use the tf.data API together with an estimator, which does
most of the work for you.
• Combine files into large files: Reading a file takes time. If you have to read
thousands of small files, this can significantly slow you down. TensorFlow
offers its own data format called TFRecord. You can also just fuse an entire
batch into a single NumPy array and save that array instead of every
example.
• Train with the tf.data.Dataset API: If you are using the TensorFlow
version of Keras, you can use the Dataset API, which optimizes data loading
and processing for you. The Dataset API is the recommended way to
load data into TensorFlow. It offers a wide range of ways to load data, for
instance, from a CSV file with tf.data.TextLineDataset, or from TFRecord
files with tf.data.TFRecordDataset.
In this example, we will use the dataset API with NumPy arrays that we have
already loaded into RAM, such as the MNIST database.
First, we create two plain datasets for data and targets:
dxtrain = tf.data.Dataset.from_tensor_slices(x_test)
dytrain = tf.data.Dataset.from_tensor_slices(y_train)
[ 359 ]
Privacy, Debugging, and Launching Your Products
dytrain = dytrain.map(apply_one_hot,num_parallel_calls=4)
We zip the data and targets into one dataset. We instruct TensorFlow to
shuffle the data when loading, keeping 200 instances in memory from which
to draw samples. Finally, we make the dataset yield batches of batch size 32:
train_data =
tf.data.Dataset.zip((dxtrain,dytrain)).shuffle(200).batch(32)
We can now fit a Keras model on this dataset just as we would fit it to a
generator:
model.fit(dataset, epochs=10, steps_per_epoch=60000 // 32)
If you have truly large datasets, the more you can parallelize, the better.
Parallelization does come with overhead costs, however, and not every problem
actually features huge datasets. In these cases, refrain from trying to do too much
in parallel and focus on slimming down your network, using CPUs and keeping all
your data in RAM if possible.
Note: You can read the full Cython documentation here: http://
cython.readthedocs.io. This section is a short introduction to
Cython. If performance is important to your application, you should
consider diving deeper.
[ 360 ]
Chapter 8
Say you have a Python function that prints out the Fibonacci series up to a specified
point. This code snippet is taken straight from the Python documentation:
from __future__ import print_function
def fib(n):
a, b = 0, 1
while b < n:
print(b, end=' ')
a, b = b, a + b
print()
Note that we have to import the print_function to make sure that print()
works in the Python 3 style. To use this snippet with Cython, save it as cython_
fib_8_7.pyx.
setup( #3
ext_modules=cythonize("cython_fib_8_7.pyx"),
)
1. The setup function is a Python function to create modules, such as the ones
you install with pip.
2. cythonize is a function to turn a pyx Python file into Cython C code.
3. We create a new model by calling setup and passing on our Cythonized code.
This will create a C file, a build file, and a compiled module. We can import this
module now by running:
import cython_fib_8_7
cython_fib_8_7.fib(1000)
This will print out the Fibonacci numbers up to 1,000. Cython also comes with
a handy debugger that shows where Cython has to fall back onto Python code,
which will slow things down. Type the following command into your Terminal:
cython -a cython_fib_8_7.pyx
[ 361 ]
Privacy, Debugging, and Launching Your Products
This will create an HTML file that looks similar to this when opened in a browser:
Cython profile
As you can see, Cython has to fall back on Python all the time in our script because
we did not specify the types of variables. By letting Cython know what data type
a variable has, we can speed up the code significantly. To define a variable with
a type, we use cdef:
from __future__ import print_function
def fib(int n):
cdef int a = 0
cdef int b = 1
while b < n:
print(b, end=' ')
a, b = b, a + b
print()
[ 362 ]
Chapter 8
Exercises
Now that we're at the end of this chapter, it's time to put what we've learned
into use. Using the knowledge that you've gained in this chapter, why not try
the following exercises?
Summary
In this chapter, you have learned a number of practical tips for debugging
and improving your model. Let's recap all of the things that we have looked at:
• Finding flaws in your data that lead to flaws in your learned model
• Using creative tricks to make your model learn more from less data
• Unit testing data in production or training to make sure standards are met
• Being mindful of privacy
• Preparing data for training and avoiding common pitfalls
• Inspecting the model and peering into the "black box"
• Finding optimal hyperparameters
• Scheduling learning rates in order to reduce overfitting
• Monitoring training progress with TensorBoard
• Deploying machine learning products and iterating on them
• Speeding up training and inference
You now have a substantial number of tools in your toolbox that will help you run
actual, practical machine learning projects and deploy them in real-life (for example,
trading) applications.
Making sure your model works before deploying it is crucial and failure to properly
scrutinize your model can cost you, your employer, or your clients millions of
dollars. For these reasons, some firms are reluctant to deploy machine learning
models into trading at all. They fear that they will never understand the models
and thus won't be able to manage them in a production environment. Hopefully, this
chapter alleviates that fear by showcasing some practical tools that can make models
understandable, generalizable, and safe to deploy.
[ 363 ]
Privacy, Debugging, and Launching Your Products
In the next chapter, we will look at a special, persistent, and dangerous problem
associated with machine learning models: bias. Statistical models tend to fit to and
amplify human biases. Financial institutions have to follow strict regulations to
prevent them from being racially or gender biased. Our focus will be to see how
we can detect and remove biases from our models in order to make them both fair
and compliant.
[ 364 ]
Fighting Bias
We like to think that machines are more rational than us: heartless silicon
applying cold logic. Thus, when computer science introduced automated decision
making into the economy, many hoped that computers would reduce prejudice and
discrimination. Yet, as we mentioned earlier when looking at mortgage applications
and ethnicity, computers are made and trained by humans, and the data that
those machines use stems from an unjust world. Simply put, if we are not careful,
our programs will amplify human biases.
While it's okay to target ads for books from a certain country to people who are also
from that country, it's usually not okay, and thanks to the ECOA, often illegal, to
deny a loan to people from a certain country. Within the financial domain, there are
much stricter rules for discrimination than those seen in book sales. This is because
decisions in the financial domain have a much more severe impact on people's lives
than those of book sales.
Equally, discrimination in this context is feature specific. For example, while it's
okay to discriminate against loan applicants based on their history of repaying loans,
it's not okay to do so based on their country of origin, unless there are sanctions
against that country or similar overarching laws in place.
[ 365 ]
Fighting Bias
The algorithms discussed in this book are feature extraction algorithms. Even
if regulated features are omitted, an algorithm might infer them from proxy features
and then discriminate based on them anyway. As an example of this, ZIP codes can
be used to predict race reasonably well in many cities in the United States. Therefore,
omitting regulated features is not enough when it comes to combating bias.
This disparate sample size can lead to worse model performance for the minority
group. As a result, this increased error is often known as a systematic error. The
model might have to overfit the majority group data so that the relationships it found
do not apply to the minority group data. Since there is little minority group data, this
is not punished as much.
Imagine you are training a credit scoring model, and the clear majority of your
data comes from people living in lower Manhattan, and a small minority of it comes
from people living in rural areas. Manhattan housing is much more expensive, so the
model might learn that you need a very high income to buy an apartment. However,
rural housing is much cheaper in comparison. Even so, because the model is largely
trained on data from Manhattan, it might deny loan applications to rural applicants
because they tend to have lower incomes than their Manhattan peers.
Aside from sample size issues, our data can be biased by itself. For example,
"raw data" does not exist. Data does not appear naturally, instead it's measured
by humans using human-made measurement protocols, which in themselves can
be biased in many different ways.
[ 366 ]
Chapter 9
Biases could include having sampling biases, such as in the Manhattan housing
example, or having measurement biases, which is when your sample might not
measure what it is intended to measure, or may even discriminate against one group.
Another bias that's possible is pre-existing social biases. These are visible in word
vectors, for instance, in Word2Vec, where the mapping from father to doctor in
latent space maps from mother to nurse. Likewise, the vector from man to computer
programmer maps from woman to homemaker. This is because sexism is encoded
within the written language of our sexist society. Until today, typically speaking
doctors have usually been men and nurses have usually been women. Likewise, tech
companies' diversity statistics reveal that far more men are computer programmers
than women, and these biases get encoded into models.
Legal perspectives
There are two doctrines in anti-discrimination law: disparate treatment, and disparate
impact. Let's take a minute to look at each of these:
After this is done, the plaintiff has the opportunity to show that the goal of
the procedure could also be achieved with a different procedure that shows
a smaller disparity.
[ 367 ]
Fighting Bias
Note: For a more in-depth overview of these topics, see Moritz Hardt's
2017 NeurIPS presentation on the topic at http://mrtz.org/
nips17/#/11.
The disparate treatment doctrine tries to achieve procedural fairness and equal
opportunity. The disparate impact doctrine aims for distributive justice and
minimized inequality in outcomes.
There is an intrinsic tension between the two doctrines, as illustrated by the Ricci
V. DeStefano case from 2009. In this case, 19 white firefighters and 1 Hispanic
firefighter sued their employer, the New Haven Fire Department. The firefighters
had all passed their test for promotion, yet their black colleagues did not score
the mark required for the promotion. Fearing a disparate impact lawsuit, the city
invalidated the test results and did not promote the firefighters. Because the evidence
for disparate impact was not strong enough, the Supreme Court of the United States
eventually ruled that the firefighters should have been promoted.
Given the complex legal and technical situation around fairness in machine learning,
we're going to dive into how we can define and quantify fairness, before using this
insight to create fairer models.
Observational fairness
Equality is often seen as a purely qualitative issue, and as such, it's often dismissed
by quantitative-minded modelers. As this section will show, equality can be seen
from a quantitative perspective, too. Consider a classifier, c, with input X, some
sensitive input, A, a target, Y and output C. Usually, we would denote the classifier
output as Yˆ , but for readability, we follow CS 294 and name it C.
Let's say that our classifier is being used to decide who gets a loan. When would
we consider this classifier to be fair and free of bias? To answer this question,
picture two demographics, group A and B, both loan applicants. Given a credit
score, our classifier must find a cutoff point. Let's look at the distribution of
applicants in this graph:
Note: The data for this example is synthetic; you can find the Excel
file used for these calculations in the GitHub repository of this book,
https://github.com/PacktPublishing/Machine-Learning-
for-Finance/blob/master/9.1_parity.xlsx.
[ 368 ]
Chapter 9
Max profits
For this exercise, we assume that a successful applicant yields a profit of $300, while
a defaulting successful applicant costs $700. The cutoff point here has been chosen to
maximize profits:
So, what can we see? We can see the following:
• In orange are applicants who would not have repaid the loan and did not get
accepted: true negatives (TNs).
• In blue are applicants who would have repaid the loan but did not get
accepted: false negatives (FNs).
• In yellow are applicants who did get the loan but did not pay it back: false
positives (FPs).
• In gray are applicants who did receive the loan and paid it back: true
positives (TPs).
As you can see, there are several issues with this choice of cutoff point. Group
B applicants need to have a better score to get a loan than Group A applicants,
indicating disparate treatment. At the same time, only around 51% of Group
A applicants get a loan but only 37% of Group B applicants do, indicating
disparate impact.
A group unaware threshold, which we can see below, would give both groups the same
minimum score:
[ 369 ]
Fighting Bias
Equal cutoff
In the preceding graph, while both groups have the same cutoff rate, Group A has
been given fewer loans. At the same time, predictions for Group A have a lower
accuracy than the predictions given for Group B. It seems that although both groups
face the same score threshold, Group A is at a disadvantage.
Demographic parity aims to achieve fairness by ensuring that both groups have the
same chance of receiving the loan. This method aims to achieve the same selection
rate for both groups, which is what impact disparity is measured by. Mathematically,
this process can be expressed as follows:
P (C = 1| A = 1) = P (C = 1| A = 0)
If we apply this rule to the same context as we used previously, we'll arrive at the
following cutoff points:
[ 370 ]
Chapter 9
While this method cannot be blamed for statistical discrimination and disparate
impact, it can be blamed for disparate treatment. In the equal pick rate graphic we
can see how Group A is given a lower threshold score; meanwhile, there are more
successful Group A applicants who default on their loans. In fact, Group A is not
profitable and gets subsidized by Group B. Accepting a worse economic outcome
to favor a certain group is also known as taste-based discrimination. It could be said
that the higher thresholds for Group B are unfair, as they have a lower FP rate.
TP parity, which is also called equal opportunity, means that both demographics
have the same TP rate. For people who can pay back the loan, the same chance of
getting a loan should exist. Mathematically, this can be expressed as follows:
P (C = 1| Y = 1, A = 1) = P (C = 1| Y = 1, A = 0)
Applied to our data, this policy looks similar to demographic parity, except that the
group cutoff point is even lower:
Equal opportunity
Accuracy parity tells us that the accuracy of predictions should be the same for both
groups. Mathematically, this can be expressed as follows:
P (C = Y | A = 1) = P (C = Y | A = 0)
[ 371 ]
Fighting Bias
The probability that the classifier is correct should be the same for the two possible
values of the sensitive variable A. When we apply this criteria to our data, we arrive
at the following output:
Equal accuracy
From the preceding diagram, the downside becomes apparent. In order to satisfy
the accuracy constraint, members of Group B are given much easier access to loans.
Therefore to solve this, trade-offs are necessary because no classifier can have
precision parity, TP parity, and FP parity unless the classifier is perfect. C = Y,
or both demographics have the same base rates:
P (Y = 1| A = 1) = P (Y = 1| A = 0)
There are many more ways to express fairness. The key takeaway, however, is that
none of them perfectly satisfies all of the fairness criteria. For any two populations
with unequal base rates, and unequal chances of repaying their loan, establishing
statistical parity requires the introduction of a treatment disparity.
This fact has led to a number of debates, with the best practice to express and
eliminate discrimination having not been agreed on yet. With that being said, even if
the perfect mathematical expression of fairness was found, it would not immediately
lead to perfectly fair systems.
Any machine learning algorithm is part of a bigger system. Inputs X are often not as
clearly defined as a different algorithm in the same system that might use different
inputs. Demographic groups A are often not clearly defined or inferred. Even the
output, C, of the classifier can often not be clearly distinguished, as many algorithms
together might perform the classification task while each algorithm is predicting
a different output, such as a credit score or a profitability estimate.
[ 372 ]
Chapter 9
Good technology is not a substitute for good policy. Blindly following an algorithm
without the opportunity for individual consideration or appeal will always lead to
unfairness. With that being said, while mathematical fairness criteria cannot solve
all the fairness issues that we face, it is surely worth trying to make machine learning
algorithms fairer, which is what the next section is about.
Training to be fair
There are multiple ways to train models to be fairer. A simple approach could be
using the different fairness measures that we have listed in the previous section as
an additional loss. However, in practice, this approach has turned out to have several
issues, such as having poor performance on the actual classification task.
An alternative approach is to use an adversarial network. Back in 2016, Louppe,
Kagan, and Cranmer published the paper Learning to Pivot with Adversarial Networks,
available at https://arxiv.org/abs/1611.01046. This paper showed how to use
an adversarial network to train a classifier to ignore a nuisance parameter, such as
a sensitive feature.
In this example, we will train a classifier to predict whether an adult makes over
$50,000 in annual income. The challenge here is to make our classifier unbiased
from the influences of race and gender, with it only focusing on features that we can
discriminate on, including their occupation and the gains they make from their capital.
To this end, we must train a classifier and an adversarial network. The adversarial
network aims to classify the sensitive attributes, a, gender and race, from the
predictions of the classifier:
The classifier aims to classify by income but also aims to fool the adversarial
network. The classifier's minimization objective formula is as follows:
[ 373 ]
Fighting Bias
To train this model fairly, we not only need data X and targets y, but also data about
the sensitive attributes, A. In the example we're going to work on, we'll be taking
data from the 1994 US census provided by the UCI repository: https://archive.
ics.uci.edu/ml/datasets/Adult.
To make loading the data easier, it has been transformed into a CSV file with column
headers. As a side note, please refer to the online version to see the data as viewing
the data would be difficult in the format of the book.
First, we load the data. The dataset contains data about people from a number of
different races, but for the simplicity of this task, we will only be focusing on white
and black people for the race attribute. To do this, we need to run the following
code:
path = '../input/adult.csv'
input_data = pd.read_csv(path, na_values="?")
input_data = input_data[input_data['race'].isin(['White',
'Black'])]
Next, we select the sensitive attributes, in this case we're focusing on race and
gender, into our sensitive dataset, A. We one-hot encode the data so that "Male"
equals one for the gender attribute and White equals one for the race attribute.
We can achieve this by running the following code:
sensitive_attribs = ['race', 'gender']
A = input_data[sensitive_attribs]
A = pd.get_dummies(A,drop_first=True)
A.columns = sensitive_attribs
Our target is the income attribute. Therefore, we need to encode >50K as 1 and
everything else as zero, which is achieved by writing this code:
y = (input_data['income'] == '>50K').astype(int)
[ 374 ]
Chapter 9
To get our training data, we firstly remove the sensitive and target attributes. Then
we fill all of the missing values and one-hot encode all of the data, as you can see in
the following code:
X = input_data.drop(labels=['income', 'race', 'gender'],axis=1)
X = X.fillna('Unknown')
X = pd.get_dummies(X, drop_first=True)
Finally, we split the data into train and test sets. As seen in the following code, we
then stratify the data to ensure that the same number of high earners are in both the
test and training data:
X_train, X_test, y_train, y_test, A_train, A_test = \
train_test_split(X, y, A, test_size=0.5,
stratify=y, random_state=7)
To ensure the data works nicely with the neural network, we're now going to scale
the data using scikit-learn's StandardScaler:
scaler = StandardScaler().fit(X_train)
X_train = pd.DataFrame(scaler.transform(X_train),
columns=X_train.columns,
index=X_train.index)
X_test = pd.DataFrame(scaler.transform(X_test),
columns=X_test.columns,
index=X_test.index)
We need a metric of how fair our model is. We are using the disparate impact
selection rule. The p_rule method calculates the share of people classified to have
over $50,000 income from both groups and then returns the ratio of selections in the
disadvantaged demographic over the ratio of selections in the advantaged group.
The goal is for the p_rule method to return at least 80% in order to meet the four-
fifths rule for both race and gender. The following code shows how this function
is only used for monitoring, and not as a loss function:
def p_rule(y_pred, a_values, threshold=0.5):
y_a_1 = y_pred[a_values == 1] > threshold if threshold else
y_pred[a_values == 1] #1
y_a_0 = y_pred[a_values == 0] > threshold if threshold else
y_pred[a_values == 0]
odds = y_a_1.mean() / y_a_0.mean() #2
return np.min([odds, 1/odds]) * 100
[ 375 ]
Fighting Bias
Let's explore this code in some more detail. As you can see from the preceding code
block, it's created with two key features, which are as follows:
To make the model setup a bit easier, we need to define the number of input features
and the number of sensitive features. This is something that is simply done by
running these two lines:
n_features=X_train.shape[1]
n_sensitive=A_train.shape[1]
Now we set up our classifier. Note how this classifier is a standard classification
neural network. It features three hidden layers, some dropout, and a final output
layer with a sigmoid activation, which occurs since this is a binary classification task.
This classifier is written in the Keras functional API.
To make sure you understand how the API works, go through the following code
example and ensure you understand why the steps are taken:
clf_inputs = Input(shape=(n_features,))
x = Dense(32, activation='relu')(clf_inputs)
x = Dropout(0.2)(x)
x = Dense(32, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(32, activation='relu')(x)
x = Dropout(0.2)(x)
outputs = Dense(1, activation='sigmoid', name='y')(x)
clf_net = Model(inputs=[clf_inputs], outputs=[outputs])
The adversarial network is a classifier with two heads: one to predict the applicant's
race from the model output, and one to predict the applicant's gender:
adv_inputs = Input(shape=(1,))
x = Dense(32, activation='relu')(adv_inputs)
x = Dense(32, activation='relu')(x)
x = Dense(32, activation='relu')(x)
out_race = Dense(1, activation='sigmoid')(x)
out_gender = Dense(1, activation='sigmoid')(x)
[ 376 ]
Chapter 9
adv_net = Model(inputs=[adv_inputs],
outputs=[out_race,out_gender])
From the preceding code, there are four key features that we should take a moment
to explore:
1. The function accepts a Keras neural network, for which the train switch
function will be created.
2. Inside the function, a second function is created. This second function accepts
a Boolean flag (True/False).
3. When called, the second function sets the network's trainability to the flag. If
False is passed, the network is not trainable. Since the layers of the network
can also be used in other networks, we ensure that each individual layer is
not trainable, too.
4. Finally, we return the function.
Using a function to create another function might seem convoluted at first, but this
allows us to create "switches" for the neural network easily. The following code snippet
shows us how to create switch functions for the classifier and the adversarial network:
trainable_clf_net = make_trainable_fn(clf_net)
trainable_adv_net = make_trainable_fn(adv_net)
To make the classifier trainable, we can use the function with the True flag:
trainable_clf_net(True)
Now we can compile our classifier. As you will see later on in this chapter, it is useful
to keep the classifier network as a separate variable from the compiled classifier with
which we make predictions:
clf = clf_net
clf.compile(loss='binary_crossentropy', optimizer='adam')
[ 377 ]
Fighting Bias
Remember that to train our classifier, we need to run its predictions through
the adversary as well as obtaining the adversary loss and applying the negative
adversary loss to the classifier. This is best done by packing the classifier and
adversary into one network.
To do this, we must first create a new model that maps from the classifier inputs
to the classifier and adversary outputs. We define the adversary output to be a
nested function of the adversarial network and the classifier network. This way,
the predictions of the classifier get immediately passed on to the adversary:
adv_out = adv_net(clf_net(clf_inputs))
We then define the classifier output to be the output of the classifier network, just
as we would for classification:
clf_out = clf_net(clf_inputs)
Then, we define the combined model to map from the classifier input, that is,
the data about an applicant, to the classifier output and adversary output:
clf_w_adv = Model(inputs=[clf_inputs],
outputs=[clf_out]+adv_out)
When training the combined model, we only want to update the weights of the
classifier, as we will train the adversary separately. We can use our switch functions
to make the classifier network trainable and the adversarial network untrainable:
trainable_clf_net(True)
trainable_adv_net(False)
The preceding expression leads to loss weights of [1.,-130,-30]. This means the
classification error has a weight of 1, the race prediction error of the adversary has
a weight of -130, and the gender prediction error of the adversary has a weight of
-30. Since the losses of the adversarial's prediction have negative weights, gradient
descent will optimize the parameters of the classifier to increase these losses.
Finally, we can compile the combined network:
clf_w_adv.compile(loss='binary_crossentropy'),
loss_weights=loss_weights,
optimizer='adam')
[ 378 ]
Chapter 9
With the classifier and combined classifier-adversarial model in place, the only thing
missing is a compiled adversarial model. To get this, we'll first define the adversarial
model to map from the classifier inputs to the outputs of the nested adversarial-
classifier model:
adv = Model(inputs=[clf_inputs],
outputs=adv_net(clf_net(clf_inputs)))
Then, when training the adversarial model, we want to optimize the weights of the
adversarial network and not of the classifier network, so we use our switch functions
to make the adversarial trainable and the classifier not trainable:
trainable_clf_net(False)
trainable_adv_net(True)
Finally, we compile the adversarial model just like we would with a regular
Keras model:
adv.compile(loss='binary_crossentropy', optimizer='adam')
With all the pieces in hand, we can now pretrain the classifier. This means we train
the classifier without any special fairness considerations:
trainable_clf_net(True)
clf.fit(X_train.values, y_train.values, epochs=10)
After we have trained the model, we can make predictions on the validation set to
evaluate both the model's fairness and accuracy:
y_pred = clf.predict(X_test)
Now we'll calculate the model's accuracy and p_rule for both gender and race. In all
calculations, we're going to use a cutoff point of 0.5:
acc = accuracy_score(y_test,(y_pred>0.5))* 100
print('Clf acc: {:.2f}'.format(acc))
out:
Clf acc: 85.44
race: 41.71%
gender: 29.41%
As you can see, the classifier achieves a respectable accuracy, 85.44%, in predicting
incomes. However, it is deeply unfair. It gives women only a 29.4% chance to make
over $50,000 than it does men.
[ 379 ]
Fighting Bias
Note: Neither gender or race was included in the features of the classifier.
Yet, the classifier discriminates strongly on them. If the features can be
inferred, dropping sensitive columns is not enough.
To get out of this mess, we will pretrain the adversarial network before training both
networks to make fair predictions. Once again, we use our switch functions to make
the classifier untrainable and the adversarial trainable:
trainable_clf_net(False)
trainable_adv_net(True)
As the distributions for race and gender in the data might be skewed, we're going to
use weighted classes to adjust for this:
class_weight_adv = compute_class_weights(A_train)
We then train the adversary to predict race and gender from the training data
through the predictions of the classifier:
adv.fit(X_train.values,
np.hsplit(A_train.values, A_train.shape[1]),
class_weight=class_weight_adv, epochs=10)
NumPy's hsplit function splits the 2D A_train matrix into two vectors that are
then used to train the two model heads.
With the classifier and adversary pretrained, we will now train the classifier to fool
the adversary in order to get better at spotting the classifier's discrimination. Before
we start, we need to do some setup. We want to train for 250 epochs, with a batch
size of 128, with two sensitive attributes:
n_iter=250
batch_size=128
n_sensitive = A_train.shape[1]
The combined network of the classifier and adversarial also needs some class
weights. The weights for the income predictions, less/more than $50,000, are
both one. For the adversarial heads of the combined model, we use the preceding
computed adversarial class weights:
class_weight_clf_w_adv = [{0:1., 1:1.}]+class_weight_adv
[ 380 ]
Chapter 9
To keep track of metrics, we set up one DataFrame for validation metrics, accuracy,
and area under the curve, as well as for the fairness metrics. The fairness metrics are
the p_rule values for race and gender:
val_metrics = pd.DataFrame()
fairness_metrics = pd.DataFrame()
Inside the main training loop, three steps are performed: training the adversarial
network, training the classifier to be fair, and printing out validation metrics. For
better explanations, all three are printed separately here.
Within the code, you will find them in the same loop, where idx is the current
iteration:
for idx in range(n_iter):
The first step is to train the adversarial network. To this end, we're going to make the
classifier untrainable, the adversarial network trainable, and then train the adversarial
network just as we did before. To do this, we need to run the following code block:
trainable_clf_net(False)
trainable_adv_net(True)
adv.fit(X_train.values,
np.hsplit(A_train.values, A_train.shape[1]),
batch_size=batch_size,
class_weight=class_weight_adv,
epochs=1, verbose=0)
Training the classifier to be a good classifier but also to fool the adversary and be fair
involves three steps. Firstly, we make the adversary untrainable and the classifier
trainable:
trainable_clf_net(True)
trainable_adv_net(False)
Finally, we train the combined adversary and classifier. Since the adversarial network
is set to not be trainable, only the classifier network will be trained. However, the loss
from the adversarial network's predictions of race and gender gets backpropagated
through the entire network, so that the classifier learns to fool the adversarial network:
clf_w_adv.train_on_batch(X_batch,
[ 381 ]
Fighting Bias
[y_batch]+\
np.hsplit(A_batch, n_sensitive),
class_weight=class_weight_clf_w_adv)
Finally, we want to keep track of progress by first making predictions on the test:
y_pred = pd.Series(clf.predict(X_test).ravel(),
index=y_test.index)
We then calculate the area under the curve (ROC AUC) and the accuracy of the
predictions, and save them in the val_metrics DataFrame:
roc_auc = roc_auc_score(y_test, y_pred)
acc = accuracy_score(y_test, (y_pred>0.5))*100
Next up, we calculate p_rule for both race and gender and save those values in the
fairness metrics:
for sensitive_attr :n A_test.columns:
fairness_metrics.loc[idx, sensitive_attr] =\
p_rule(y_pred,A_test[sensitive_attr])
If we plot both the fairness and validation metrics, we'll arrive at the following plot:
[ 382 ]
Chapter 9
As you can see, the fairness scores of the classifier steadily increase with training.
After about 150 epochs, the classifier satisfies the four-fifths rule. At the same time,
the p-values are well over 90%. This increase in fairness comes at only a small
decrease in accuracy and area under the curve. The classifier trained in this manner
is clearly a fairer classifier with similar performance, and is thus preferred over
a classifier trained without fairness criteria.
The pivot approach to fair machine learning has a number of advantages. Yet, it
cannot rule out unfairness entirely. What if, for example, there was a group that the
classifier discriminates against that we did not think of yet? What if it discriminates
on treatment, instead of impact? To make sure our models are not biased, we need
more technical and social tools, namely interpretability, causality, and diverse development
teams.
In the next section, we'll discuss how to train machine learning models that learn
causal relationships, instead of just statistical associations.
Causal learning
This book is by and large a book about statistical learning. Given data X and targets
Y, we aim to estimate p ( y | x) , the distribution of target values given certain data
points. Statistical learning allows us to create a number of great models with useful
applications, but it doesn't allow us to claim that X being x caused Y to be y.
Judea Pearl, a famous computer scientist, invented a notation for causal models
called do-calculus; we are interested in p ( y | do ( p)) , which is the probability of
someone behaving recklessly after we manipulated P to be p. In a causal notation,
X usually stands for observed features, while P stands for the policy features that
we can manipulate. This notation can be a bit confusing, as p now expresses both
a probability and a policy. Yet, it is important to distinguish between observed and
influenced features. So, if you see do ( p ) , p is a feature that is influenced, and if you
see p (..) , p is a probability function.
So, the formula p ( y | x) expresses the statistical relationship that insurance holders
are more reckless on average. This is what supervised models learn. p ( y | do ( p), x)
expresses the causal relationship that people who get insurance become more
reckless because they are insured.
[ 383 ]
Fighting Bias
Causal models are a great tool for fair learning. If we only build our models in
a causal way, then we'll avoid most of the statistical discrimination that occurs in
statistical models. Do females statistically earn less than males? Yes. Do females
earn less because they are females and females are somehow undeserving of high
salaries? No. Instead, the earnings difference is caused by other factors, such as
different jobs being offered to males and females, discrimination in the workplace,
cultural stereotypes, and so on.
That does not mean we have to throw statistical models out of the window. They
are great for the many cases where causality is not as much of an important factor
and where we do not intend to set the values of X. For instance, if we are creating
a natural language model, then we are not interested in whether the occurrence of
a word caused the sentence to be about a certain topic. Knowing that the topic and
the word are related is enough to make predictions about the content of the text.
[ 384 ]
Chapter 9
Instrument variables
To distinguish the influence on policy and response, we need access to an
instrument, Z. An instrument is a variable that influences the policy, but nothing
else. The reinsurance cost, for example, could prompt the insurance company to
give out fewer insurance policies. This relationship can be seen in the flowchart
below, where the relationship has been mapped:
Causal flowchart
The field of econometrics already has a built a method to work with these kinds
of situations called instrumental variables two-stage least squares (IV2SLS, or just
2SLS). In a nutshell, 2SLS first fits a linear regression model between the instrument, z,
and the policy, p, which in econometrics called the endogenous or treatment variable.
From this linear regression, it then estimates an "adjusted treatment variable," which
is the treatment variable as it can be explained by the instrument. The idea is that
this adjustment removes the influence of all other factors on the treatment. A second
linear regression model then creates a linear model mapping from the features, x,
and the adjusted treatment variable, p̂ , to the outcome, y.
In the following diagram, you can see an overview of how 2SLS works:
IV2SLS
[ 385 ]
Fighting Bias
2SLS is probably what the insurance company in our case would use since it is
an established method. We won't go into details here, beyond giving you a brief
overview of how to use 2SLS in Python. The linear model package in Python
features an easy way to run 2SLS.
If you have data X, y, P, and Z, you can run a 2SLS regression as follows:
from linearmodels.iv import IV2SLS
iv = IV2SLS(dependent=y,
exog=X,
endog=P],
instruments=Z).fit(cov_type='unadjusted')
y = g ( p, x )
p = f ( x, z )
Given these two functions, the following identity holds, if the confounding variable
has a mean of zero overall features:
E [ y | x, z ] = E g ( p, x) | x, z = ∫ g ( p, x) dF ( p | x, z )
[ 386 ]
Chapter 9
This means that if we can reliably estimate the function, g, and distribution, F, we can
make causal statements about the effects of policy p . If we have data about the actual
outcome, y, features x, policy p , and instrument z, we can optimize the following:
( 2
min ∑ yt − ∫ g ( p, xt ) dF ( p x, z )
g ∈G
t =1
The preceding function is the squared error between the predicted outcome using
the prediction function, g, and the actual outcome, y.
Notice the similarity to 2SLS. In 2SLS, we estimated F and g with two separate linear
regressions. For the more complex functions, we can also estimate them with two
separate neural networks. Back in 2017, Jason Hartfort and others presented just
such an approach with their paper, Deep IV: A Flexible Approach for Counterfactual
Prediction, - available at: http://proceedings.mlr.press/v70/hartford17a/
hartford17a.pdf - the overview of which you can see in the following diagram:
Deep IV
The idea of Deep IV is to first train a neural network to express a distribution, F(z,x),
which describes the distribution of policies given certain features, x, and instrument
values, z. A second neural network is predicting the response, y, from the estimated
policy distribution and features. Deep IV's advantage is that it can learn complex,
non-linear relationships from complex data, such as text.
The authors of the Deep IV paper have also published a custom Keras model which
is used for handling sampling and learning from a distribution part, which you can
find on GitHub: https://github.com/jhartford/DeepIV.
[ 387 ]
Fighting Bias
While their code is too long to be discussed in depth here, it is interesting to think
about what the source of our causal claim is, both in Deep IV and IV2SLS. In our
insurance case, we assumed that either having or not having an insurance would
influence behavior, not the other way around. We never showed or tested the truth
behind this direction of causality.
With these two limitations in mind, causal inference becomes a great tool and an
active area of research from which we can hope to see great results in the future.
In the best case, your discrimination-sensitive models would only contain causal
variables. In practice, this is usually not possible. However, keeping the difference
between statistical correlation in mind, as expressed by standard statistical models
and causation, can help you avoid statistical biases and wrong associations.
A final, more technical, method to reduce unfairness is to peek inside the model to
ensure it is fair. We already looked at interpretability in the last chapter, mostly to
debug data and spot overfitting, but now, we will give it another look, this time to
justify the model's predictions.
In this section, we will use a slightly more sophisticated method called SHAP (SHapley
Additive exPlanation). SHAP combines several different explanation approaches into
one neat method. This method lets us generate explanations for individual predictions
as well as for entire datasets in order to understand the model better.
[ 388 ]
Chapter 9
Tip: The example code given here is from the SHAP example notebooks.
You can find a slightly extended version of the notebook on Kaggle:
https://www.kaggle.com/jannesklaas/explaining-income-
classification-with-keras
SHAP combines seven model interpretation methods, those being LIME, Shapley
sampling values, DeepLIFT, Quantitative Input Influence (QII), layer-wise
relevance propagation, Shapley regression values, and a tree interpreter that has
two modules: a model-agnostic KernelExplainer and a TreeExplainer module
specifically for tree-based methods such as XGBoost.
The mathematics of how and when the interpreters are used is not terribly relevant
for using SHAP. In a nutshell, given a function, f, expressed through a neural
network, for instance, and a data point, x, SHAP compares f ( x) to f ( z ) where E f ( z)
is the "expected normal output" generated for a larger sample. SHAP will then
create smaller models, similar to LIME, to see which features explain the difference
between f ( x) and E f ( z) .
[ 389 ]
Fighting Bias
Next, we need to calculate the SHAP values indicating the importance of different
features for a single example. We let SHAP create 500 permutations of each sample
from z so that SHAP has a total of 50,000 examples to compare the one example to:
shap_values = explainer.shap_values(X.iloc[350,:], nsamples=500)
Finally, we can plot the influence of the features with SHAP's own plotting tool.
This time, we provide a row from X_display, not X. X_display, which contains the
unscaled values and is only used for annotation of the plot to make it easier to read:
shap.force_plot(explainer.expected_value, shap_values)
If you look at the preceding plot, the predictions of the model seem, by and large,
reasonable. The model gives the applicant a high chance of having a high income
because they have a master's degree, and because they're an executive manager who
works 65 hours a week. The applicant could have an even higher expected income score
were it not for a capital loss. Likewise, the model seems to take the fact that the applicant
is married as a big factor of a high income. In fact, in our example, it seems that marriage
is more important than either the long hours or the job title.
Our model also has some problems that become clear once we calculate and plot the
SHAP values of another applicant:
shap_values = explainer.shap_values(X.iloc[167,:], nsamples=500)
shap.force_plot(explainer.expected_value, shap_values)
The following outputted graph is then shown. This also shows some of the problems
that we've encountered:
[ 390 ]
Chapter 9
In this example, the applicant also has a good education, and works 48 hours a week
in the technology industry, but the model gives her a much lower chance of having a
high income because of the fact that she's a female, an Asian-Pacific islander who has
never been married and has no other family relationship. A loan rejection on these
grounds is a lawsuit waiting to happen as per the ECOA act.
The two individual cases that we just looked at might have been unfortunate glitches
by the model. It might have overfitted to some strange combination that gave
an undue importance to marriage. To investigate whether our model is biased, we
should investigate a number of different predictions. Fortunately for us, the SHAP
library has a number of tools that can do just that.
Then, we can plot a forced plot for all of these values as well:
shap.force_plot(explainer.expected_value, shap_values)
Again, this code produces a SHAP dataset graph, which we can see in the following
graphic:
SHAP dataset
The preceding plot shows 230 rows of the dataset, grouped by similarity with the
forces of each feature that matter to them. In your live version, if you move the
mouse over the graph, you'll be able to read the features and their values.
[ 391 ]
Fighting Bias
By exploring this graph, you can get an idea of what kind of people the model
classifies as either high or low earners. On the very left, for example, you'll see most
people with low education who work as cleaners. The big red block between 40 and
60 are mostly highly educated people who work a high number of hours.
To further examine the impact of marital status, you can change what SHAP displays
on the y-axis. Let's look at the impact of marriage:
As you can see in this chart, marriage status either strongly positively or negatively
impacts people from different groups. If you move your mouse over the chart, you
can see that the positive influences all stem from civic marriages.
Using a summary plot, we can see which features matter the most to our model:
shap.summary_plot(shap_values, X.iloc[100:330,:])
This code then outputs the final summary plot graph, which we can see below:
[ 392 ]
Chapter 9
As you can see, education is the most important influence on our model. It also has
the widest spread of influence. Low education levels really drag predictions down,
while strong education levels really boost predictions up. Marital status is the second
most important predictor. Interestingly, though, capital losses are important to the
model, but capital gains are not.
[ 393 ]
Fighting Bias
To dig deeper into the effects of marriage, we have one more tool at our disposal, a
dependence plot, which can show the SHAP values of an individual feature together
with a feature for which SHAP suspects high interaction. With the following code
snippet, we can inspect the effect of marriage on our model's predictions:
shap.dependence_plot("marital-status",
shap_values,
X.iloc[100:330,:],
display_features=X_display.iloc[100:330,:])
As a result of running this code, we can now see a visualized representation of the
effect of marriage in the following graph:
[ 394 ]
Chapter 9
As you can see, Married-civ-spouse, the census code for a civilian marriage with no
partner in the armed forces, stands out with a positive influence on model outcomes.
Meanwhile, every other type of arrangement has slightly negative scores, especially
never married.
Statistically, rich people tend to stay married for longer, and younger people
are more likely to have never been married. Our model correctly correlated that
marriage goes hand in hand with high income, but not because marriage causes high
income. The model is correct in making the correlation, but it would be false to make
decisions based on the model. By selecting, we effectively manipulate the features on
which we select. We are no longer interested in just p ( y | x) , but in p ( y | do ( p)) .
Unfairness is equally complex. We've already discussed the two general definitions
of unfairness, disparate impact and disparate treatment. Disparate treatment can occur
against any combination of features (age, gender, race, nationality, income, and so
on), often in complex and non-linear ways. This section examines Richard Cook's
1998 paper, How complex systems fail - available at https://web.mit.edu/2.75/
resources/random/How%20Complex%20Systems%20Fail.pdf - which looks at
how complex machine learning-driven systems fail to be fair. Cook lists 18 points,
some of which will be discussed in the following sections.
[ 396 ]
Chapter 9
Exercises
In this chapter, you have learned a lot about both the technical and non-technical
considerations of fairness in machine learning. These exercises will help you think
much more deeply about the topic:
• Think about the organization you work for. How is fairness incorporated
in your organization? What works well and what could be improved?
• Revisit any of the models developed in this book. Are they fair? How would
you test them for fairness?
• Fairness is only one of the many complex issues large models can have. Can
you think of an issue in your area of work that could be tackled with the tools
discussed in this chapter?
Summary
In this chapter, you have learned about fairness in machine learning in different
aspects. First, we discussed legal definitions of fairness and quantitative ways to
measure these definitions. We then discussed technical methods to train models
to meet fairness criteria. We also discussed causal models. We learned about SHAP
as a powerful tool to interpret models and find unfairness in a model. Finally, we
learned how fairness is a complex systems issue and how lessons from complex
systems management can be applied to make models fair.
[ 398 ]
Chapter 9
There is no guarantee that following all the steps outlined here will make your model
fair, but these tools vastly increase your chances of creating a fair model. Remember
that models in finance operate in high-stakes environments and need to meet many
regulatory demands. If you fail to do so, damage could be severe.
In the next, and final, chapter of this book, we will be looking at probabilistic
programming and Bayesian inference.
[ 399 ]
Bayesian Inference and
Probabilistic Programming
Mathematics is a big space of which humans so far have only charted a small amount.
We know of countless areas in mathematics that we would like to visit, but that are not
tractable computationally.
In the final chapter of this book, you'll learn about the following:
This book has largely covered deep learning and its applications in the finance
industry. As we've witnessed, deep learning has been made practical through
modern computing power, but it is not the only technique benefiting from this
large increase in power.
[ 401 ]
Bayesian Inference and Probabilistic Programming
Both Bayesian inference and probabilistic programming are two up and coming
techniques whose recent progress is powered by the increase in computing power.
While the advances in the field have received significantly less press coverage than
deep learning, they might be even more useful to the financial practitioner.
Bayesian models are interpretable and can express uncertainty naturally. They
are less "black box," instead making the modeler's assumptions more explicit.
This example is similar to the one given in the 2015 book, Bayesian Methods for
Hackers: Probabilistic Programming and Bayesian Inference, written by Cameron
Davidson-Pilon. However, in our case, this is adapted to a financial context
and rewritten so that the mathematical concepts intuitively arise from the code.
Note: You can view the example at the following link: http://
camdavidsonpilon.github.io/Probabilistic-Programming-
and-Bayesian-Methods-for-Hackers/.
Let's imagine that you have a security that can either pay $1 or, alternatively,
nothing. The payoff depends on a two-step process. With a 50% probability, the
payoff is random, with a 50% chance of getting $1 and a 50% chance of making
nothing. The 50% chance of getting the dollar is the true payoff probability (TPP), x.
[ 402 ]
Chapter 10
Payoff scheme
You are interested in finding out what the true payoff ratio is, as it will inform your
trading strategy. In our case, your boss allows you to buy 100 units of securities. You
do, and 54 of the 100 securities pay you a dollar.
But what is the actual TPP? In this case, there is an analytical solution to calculate the
most likely TPP, but we will be using a computational method that also works for
more complicated cases.
Flat prior
The variable x represents the TPP. We randomly sample 100 truth values, which are 1
if you had gotten the dollar under the true payoff, and 0 if otherwise. We also sample
the two random choices at Start and Random Payoff in the preceding scheme. It
is computationally more efficient to sample the random outcomes in one go for all
trials, even though they are not all needed.
[ 403 ]
Bayesian Inference and Probabilistic Programming
Finally, we sum up the payoffs and divide them by the number of securities in our
simulation in order to obtain the share of payoffs in the simulation.
The following code snippet runs one simulation. It's important, though, to make
sure that you understand how the computations follow from our securities structure:
def run_sim(x):
truth = np.random.uniform(size=100) < x
first_random = np.random.randint(2,size=100)
second_random = np.random.randint(2,size=100)
res = np.sum(first_random*truth +
(1-first_random)*second_random)/100
return res
Next, we would like to try out a number of possible TPPs. So, in our case, we'll
sample a candidate TPP and run the simulation with the candidate probability. If the
simulation outputs the same payoff as we observed in real life, then our candidate
is a real possibility.
The following sample method returns real possibilities, or None if the candidate
it tried out was not suitable:
def sample(data = 0.54):
x = np.random.uniform()
if run_sim(x) == data:
return x
As we have to sample a number of possible TPPs, it's only natural that we want to
speed this process up. To do this, we can use a library called JobLib, which will help
with parallel execution.
To do this, we need to import the Parallel class, which will help to run loops in
parallel, and the delayed method, which helps to execute functions in order inside
the parallel loop. We can import them by running the following:
from JobLib import Parallel, delayed
[ 404 ]
Chapter 10
The details are not relevant for this example, but the Parallel(n_jobs=-1)
method makes the job run with as many parallel executions as there are CPUs on
the machine. For example, delayed(sample)() for i in range(100000) runs
the sample method 100,000 times.
We obtain a Python list, t, which we turn into a NumPy array. As you can see in
the following code snippet, about 98% of the array are None values. That means that
98% of the values the sampler tried out for x did not yield results matching our data:
t = Parallel(n_jobs=-1)(delayed(sample)() for i in range(100000))
t = np.array(t,dtype=float)
share = np.sum(np.isnan(t))/len(t)*100
print(f'{share:.2f}% are throwaways')
Therefore, we'll now throw away all of the None values, leaving us with the possible
values for x:
t_flat = t[~np.isnan(t)]
plt.hist(t_flat, bins=30,density=True)
plt.title('Distribution of possible TPPs')
plt.xlim(0,1);
[ 405 ]
Bayesian Inference and Probabilistic Programming
As you can see, there is a distribution of possible TPPs. What this graph shows us
is that the most likely TPP is somewhere around 50% to 60%; though other values
are possible, they are somewhat less likely.
What you've just seen is one of the big advantages of Bayesian methods. All of the
estimates come in distributions, for which we can then calculate confidence intervals,
or credibility intervals, as they are known in Bayesian terminology.
This allows us to be more precise about how sure we are about things and what
other values parameters in our model could have. Relating it back to our interest
in finance, with financial applications, where millions are staked on the outputs
of models, it becomes very advantageous to quantify such uncertainty.
<50% prior
At this point, you are able to take your results to your boss, who is a domain
expert on the securities that you are trading. He looks at your analysis and shakes
his head saying, "The TPP cannot be more than 0.5." He explains, "From the underlying
business, it's physically impossible to do more than that."
So, how can you incorporate this fact into your simulation analysis? Well, the
straightforward solution is to only try out candidate TPPs from 0 to 0.5. All you
have to do is to limit the space you sample the candidate value of x, which can be
achieved by running the following code:
def sample(data = 0.54):
x = np.random.uniform(low=0,high=0.5)
if run_sim(x) == data:
return x
t_cut = t[~np.isnan(t)]
plt.hist(t_cut, bins=15,density=True)
plt.title('Distribution of possible TPPs')
plt.xlim(0,1);
[ 406 ]
Chapter 10
The first time around, you believed that all TPPs between 0 and 100% were equally
likely before seeing any data. This is called a flat prior, as the distribution of values
is the same for all values and is therefore flat. The second time, you believed that the
TPPs had to be below 50%.
The distribution expressing your beliefs about x before seeing the data is called the
prior distribution, P(TPP), or just prior. The distribution of the possible values of x
that we obtained from simulation, that is, after seeing data D, is called the posterior
distribution, P (TPP D ) , or just posterior.
The following plots show samples from prior and posterior for the first and second
rounds. The first plot shows the results with a flat posterior:
flat_prior = np.random.uniform(size=1000000)
plt.hist(flat_prior,bins=10,density=True, label='Prior')
plt.hist(t_flat, bins=30,density=True, label='Posterior')
plt.title('Distribution of $x$ with no assumptions')
plt.legend()
plt.xlim(0,1);
[ 407 ]
Bayesian Inference and Probabilistic Programming
The next plot shows the output of our sampler with a <50% prior:
cut_prior = np.random.uniform(low=0,high=0.5,size=1000000)
plt.hist(cut_prior,bins=10,density=True, label='Prior')
plt.hist(t_cut, bins=15,density=True, label='Posterior')
plt.title('Distribution of $x$ assuming TPP <50%')
plt.legend()
plt.xlim(0,1);
While it's still the same sampler, you can see that the outcome is quite different:
[ 408 ]
Chapter 10
Have you noticed anything curious? The posterior values of the second round are
roughly equal to the posterior values of the first round, but here they are cut off
at 0.5. This is because the second round prior is 0 for values above 0.5 and 1 for
everywhere else.
As we only keep simulation results that match the data, the number of kept
simulation results shown in the histogram reflects the probability of running
a simulation that yields the observed data D for a given TPP, C, P ( D TPP ) .
The posterior probabilities, P ( C D ) , that we obtain from our simulations are equal
to the probability that we observe the data when trying out a given TPP, P ( D TPP ) ,
times the probability, P(TPP).
Mathematically, this is represented as follows:
P (TPP D ) = P ( D TPP ) P (TPP )
When the data is naturally obtained, such as through a face-to-face meeting, then
we might need to account for biases in our data collection method. Most of the time,
we do not have to worry about this and can simply leave it out, but sometimes the
measurement can amplify certain outcomes.
To mitigate this, we'll divide by the data distribution, P ( D ) , as a final addon to our
posterior formula and arrive at the following formula:
P ( D TPP ) P (TPP )
P (TPP D ) =
P ( D)
As you can see, it's the Bayes formula! When running our simulation, we are
sampling from the posterior. So, why can't we just use the Bayes formula to
calculate the posterior? The simple answer is because evaluating P ( D TPP )
requires integrating over TPP, which is intractable. Our simulation method
is, as an alternative, a simple and convenient workaround.
Note: The first round prior (all TPPs are equally likely) is called a
"flat prior" because we make no assumptions about the distributions
of values. In this case, the Bayesian posterior is equal to the maximum
likelihood estimate.
Eventually, there will be too many possible parameter combinations that have no
chance of generating our data. Therefore, we need to guide our search and sample
parameters with higher posterior probabilities more often.
The approach of a guided, but still random, sampling is called the "Markov Chain
Monte Carlo algorithm". The "Monte Carlo" component means that randomness
and simulation are involved, whereas the "Markov Chain" means that we move
over the parameter space under certain probabilities.
The following diagram shows the basic workings of the Markov Chain Monte Carlo
algorithm:
What the image shows is that we are on a "random walk" in which we more or less
randomly go over different parameter values. However, we don't move entirely
randomly, but instead prefer parameter values that have high posterior probabilities.
1. Propose a new parameter value, xcand , from our current parameter value, x.
2. Estimate the posterior probability of xcand , π ( xcand ) . We can use the Bayes
rule for this.
[ 410 ]
Chapter 10
π ( xcand )
α = min 1,
π ( x)
4. Move to the new parameter value with probability α .
First, we need to propose a new Xc. This has to be dependent on the previous value
of x since we do not want a blind random search, but a more refined random walk. In
this case, we will sample xcand from a normal distribution with mean x and a standard
deviation of 0.1.
It's also possible to sample from other distributions or with other standard
deviations, as long as xcand is related to x:
def propose(x):
return np.random.randn() * 0.1 + x
In the first section, by sampling from the prior and then running the simulation, we
sampled directly from the posterior. As we are now sampling through our proposed
method, we are no longer sampling from the posterior directly. Therefore, to
calculate the posterior probability, we'll use the Bayes rule.
[ 411 ]
Bayesian Inference and Probabilistic Programming
For starters, we will use a flat prior again; each TPP is equally likely:
def prior(x):
return 1 #Flat prior
The posterior probability of a parameter value, x, is the likelihood times the prior:
def posterior(x):
return likelihood(x) * prior(x)
Now we are ready to put it all together into the Metropolis-Hastings MCMC
algorithm!
First, we need to set some initial value for x. To make the algorithm find likely
values quickly, it is sensible to initialize it at the maximum likelihood value or some
estimate that we deem likely. We also need to compute the posterior probability of
this initial value, which we can do by running the following code:
x = 0.5
pi_x = posterior(x)
Likewise, we need to keep track of all of the values sampled in a trace. Purely for
exhibition purposes, we will also keep track of the posterior probabilities. To do
this, we're going to run the following:
trace = [x]
pi_trace = [pi_x]
Now we get to the main loop. However, before we do, it's important to remember
that the algorithm consists of four steps:
[ 412 ]
Chapter 10
x_cand = propose(x)
pi_x_cand = posterior(x_cand)
u = np.random.uniform()
if i % 10 == 0:
print(f'Epoch {i}, X = {x:.2f}, pi = {pi_x:.2f}')
After running this algorithm for a number of epochs, we end up with a distribution
of possible cheater shares with payoffs. As we've done before, we can simply run the
following code to visualize this:
plt.hist(trace,bins=30)
plt.title('Metropolis Hastings Outcome')
plt.xlim(0,1);
Once we've run the previous code, we'll receive this graph as the output:
[ 413 ]
Bayesian Inference and Probabilistic Programming
By viewing the trace over time, it shows how the algorithm moves randomly
but centers around highly likely values:
plt.plot(trace)
plt.title('MH Trace');
We will then get an output, in the form of a chart, which shows us the trace of the
Metropolis Hasings (MH) sampler:
For a better understanding, we can plot the posterior probabilities over the tried out
values:
plt.scatter(x=trace,y=pi_trace)
plt.xlabel('Proposed X')
plt.ylabel('Posterior Probability')
plt.title('X vs Pi');
After successful executing the code, we'll then get the following chart as an output:
[ 414 ]
Chapter 10
Metropolis-Hastings MCMC
To demonstrate the power and flexibility of PyMC3, we are going to use it
for a classic econometrics task, but we will put a Bayesian spin on it.
Stock prices and other financial asset prices fluctuate, and the variance of daily
returns is called volatility. Volatility is a commonly used risk measure, so it's quite
important to measure it accurately.
In the example we're looking at, we are interested in the closing prices, so we need
to extract the closing prices from the dataset. The dataset shows new data first, so
we need to invert it, which we achieve with the following code:
close = pd.Series(df.Close.values,index=pd.DatetimeIndex(df.Date))
close = close[::-1]
When plotting the closing prices, which we do in the following code, we see, through
the outputted graphic, a familiar plot:
close.plot(title='S&P 500 From Inception');
SP500
[ 415 ]
Bayesian Inference and Probabilistic Programming
The dataset contains the S&P since its inception, which for us is a bit too much,
so in our case, we're going to cut it off at 1990. We can specify this date by running
the following:
close = close['1990-01-01':]
As we are interested in the returns, we need to compute the price differences. We can
use np.diff to get daily price differences. We are going to package the whole thing
into a pandas series for easier plotting:
returns = pd.Series(np.diff(close.values),index=close.index[1:])
returns.plot();
[ 416 ]
Chapter 10
Now the fun with PyMC3 begins. PyMC3 includes some special distributions for
dealing with time series, such as a random walk. This is exactly the right thing to
use when we want to model stock prices.
Firstly, we need to import PyMC3 and its tool for time series, the random walk class:
import pymc3 as pm
from pymc3.distributions.timeseries import GaussianRandomWalk
Then lastly, we need to set up the model. We can achieve this by running the
following code:
with pm.Model() as model:
step_size = pm.Exponential('sigma', 50.) #1
s = GaussianRandomWalk('s', sd=step_size, #2
shape=len(returns))
nu = pm.Exponential('nu', .1) #3
r = pm.StudentT('r', nu=nu, #4
lam=pm.math.exp(-2*s),
observed=returns.values)
Let's now look at the commands we just executed in order to set up the model.
As you can see, it consists of four key elements:
The standard sampler for PyMC3 is not Metropolis Hastings, but the No-U-Turn
Sampler (NUTS). PyMC3 will default to NUTS if we specify no sampler and just
call sample.
[ 417 ]
Bayesian Inference and Probabilistic Programming
To make the sampling run smoothly here, we need to specify a relatively high
amount of tune samples. Those are samples that the sampler will draw from in
order to find a good starting point, and that will not be part of the posterior, similar
to the burned samples before.
We also need to tell NUTS to be lenient when accepting values by setting a high
target_accept value. We can achieve this by running the following:
with model:
trace = pm.sample(tune=2000,
nuts_kwargs=dict(target_accept=.9))
PyMC3 has a nice utility that we can use to visualize the outcomes of sampling. We
are interested in the standard deviation of the volatility random walk, σ , as well as
the degrees of freedom of the StudentT distribution from which the actual returns
are drawn.
As we ran two chains in parallel, you can see that we obtained two different output
distributions. If we had run the sampler for longer, those two outcomes would
converge. We can obtain a better estimate by averaging them, which is what PyMC3
does for predictions. For instance, let's now try that with the following code:
pm.traceplot(trace, varnames=['sigma', 'nu']);
TracePlot
With the result of that code being shown in the following charts:
Results overview of the PyMC3 sampler. On the left, you can see the distributions
produced by the two sampler chains. On the right, you can see their traces.
In the final step, we can show how stochastic volatility has behaved over time.
You can see how it nicely aligns with volatile periods such as the 2008 financial
crisis. You can also see that there are periods when the model is more or less certain
about volatility:
[ 418 ]
Chapter 10
plt.plot(returns.values)
plt.plot(np.exp(trace[s].T), 'r', alpha=.03);
plt.xlabel('time')
plt.ylabel('returns')
plt.legend(['S&P500', 'Stochastic Vol.']);
As we can see, the output of that code will return the chart that we see below:
There are a large number of applications that can be modeled well with such relatively
small Bayesian models. The main advantage is that the models are easy to interpret
and can express uncertainty well. Probabilistic programming aligns well with the
"storytelling" approach to data science, as the story is clearly expressed in the model.
In the next section, we will move from shallow probabilistic programming to deep
probabilistic programming.
Summary
In this chapter, you got a brief overview of modern Bayesian machine learning
and its applications in finance. We've only touched upon this as it is a very active field
of research from which we can expect many breakthroughs in the near future. It will be
exciting to observe its development and bring its applications into production.
Looking back at this chapter, we should feel confident in understanding the following:
[ 420 ]
Chapter 10
Notice how everything you have learned here transfers to bigger models as well,
such as the deep neural networks that we've discussed throughout the entirety of the
book. The sampling process is still a bit slow for very large models, but researchers
are actively working on making it faster, and what you've learned is a great
foundation for the future.
Farewell
And thus, we close the last chapter of our journey, and I say goodbye to you, dear
reader. Let's look back at the table of contents that we were met with at the start
of our journey.
Over the past 10 chapters, we've covered a whole lot, including the following:
In each chapter, we created a large bag of practical tips and tricks that you can use.
This will allow you to build state-of-the-art systems that will change the financial
industry.
Yet, in many ways we have only scratched the surface. Each of the chapter topics
merit their own book, and even that would not adequately cover everything that
could be said about machine learning in finance.
I leave you with this thought: Machine learning in finance is an exciting field in
which there is still much to uncover, so onward dear reader; there are models to
be trained, data to be analyzed, and inferences to be made!
[ 421 ]
Bayesian Inference and Probabilistic Programming
Further reading
You made it to the end of the book! What are you going to do now? Read more
books! Machine learning, and in particular, deep learning, is a fast-moving field, so
any reading list risks being outdated by the time you read it. However, the following
list aims to show you the most relevant books that have a safety net of remaining
relevant over the coming years.
Wes is the original creator of pandas, a popular Python data-handling tool that
we saw in Chapter 2, Applying Machine Learning to Structured Data. pandas is a
core component of any data science workflow in Python and will remain so for
the foreseeable future. Investing in sound knowledge of the tools he presents is
definitely worth your time.
The "bible" of statistical machine learning, containing good explanations of all the
important concepts of statistical learning. This book is best used as a lookup book
whenever you need some in-depth information on one concept.
Gareth James, Daniela Witten, Trevor Hastie, and Robert Tibshirani, Introduction
to Statistical Learning, https://www-bcf.usc.edu/~gareth/ISL/.
[ 422 ]
Chapter 10
While this book is very praxis-oriented, Deep Learning is more focused on the
theory behind deep learning. It covers a broad range of topics and derives practical
applications from theoretical concepts.
Reinforcement learning
Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction,
http://incompleteideas.net/book/the-book-2nd.html.
The standard work of reinforcement learning discusses all major algorithms in depth.
The focus is less on flashy results and more on the reasoning behind and derivation
of reinforcement learning algorithms.
This book covers machine learning techniques from a probabilistic and much more
Bayesian perspective. It's a very good guide if you want to think about machine
learning differently.
This is probably the only probabilistic programming book that focuses on practical
applications. Not only is it free and open source, it also gets frequent updates with
new libraries and tools so that it always stays relevant.
[ 423 ]
Other Books
You May Enjoy
If you enjoyed this book, you may be interested in these other books by Packt:
Tokenomics
ISBN: 978-1-78913-632-6
Imran Bashir
ISBN: 978-1-78883-904-4
[ 426 ]
Other Books You May Enjoy
ISBN: 978-1-78712-593-3
●● Understand the key frameworks in data science, machine learning, and deep
learning
●● Harness the power of the latest Python open source libraries in machine learning
●● Explore machine learning techniques using challenging real-world data
●● Master deep neural network implementation using the TensorFlow library
●● Learn the mechanics of classification algorithms to implement the best tool
for the job
●● Predict continuous target outcomes using regression analysis
●● Uncover hidden patterns and structures in data with clustering
●● Delve deeper into textual and social media data using sentiment analysis
[ 427 ]
Other Books You May Enjoy
[ 428 ]
Index
Symbols autoencoders
about 228, 229
2010 Flash Crash use case 251 for credit cards 233-236
for MNIST 229-232
A Automatic Differentiation Variational
Inference (AVI) 420
activation function 18 automatic teller banking machine 1
active learning autoregression 139
about 266 Autoregressive Integrated Moving
generative models, using 269 Average (AIRMA) 139-142
labeling budgets, using 266 auto-sklearn
leverage machines, for human reference 318
labeling 267, 268 AutoWEKA
pseudo labeling, for unlabelled data 268 reference 318
activity_regularizer 94 AWS deep learning AMI
adam (adaptive momentum estimation) 91 reference 13
adam optimizer 91, 92 using 13
advantage actor-critic (A2C) model
about 292, 293
pendulum, balancing 294-305
B
trading 306-310 backtesting
aggregate global feature statistics 118-122 about 136, 137
alternative data 2 biases 136
Amazon Mechanical Turk (MTurk) 176 bag of words classification 191, 192
Anaconda batchnorm 97-99
reference 12 Bayesian deep learning 161-164
anti-discrimination law Bayesian inference
disparate impact 367 < 50% prior 406
disparate treatment 367 about 402, 403
approaches, beyond image classification flat prior 403, 405
about 108 Markov Chain Monte Carlo 409-414
bounding box prediction 111, 112 posterior 407, 409
facial recognition 109, 110 prior 407, 409
asynchronous advantage probabilistic programming, migrating to deep
actor-critic (A3C) 294 probabilistic programming 420
attention mechanism 209-212 stochastic volatility example 415-419
autocorrelation 131-133 behavioral economics 318
Index
[ 430 ]
Index
[ 431 ]
Index
[ 432 ]
Index
Q S
Q-function 280 sampling biases 367
Q-learning Seaborn 72
about 279 semi-supervised generative adversarial
exploration 284 network (SGAN)
used, for conversion of RL into supervised about 269
learning 281 reference 270
model, defining 284 used, for fraud detection 269-276
qualitative rationale 51 semi-supervised learning 266
Seq2Seq models
R about 213
architecture overview 213, 214
recurrent dropout 160 data 214
recurrent neural networks (RNN) 156 inference models, creating 221, 222
regex module translations, creating 222-225
using 187 SHAP (SHapley Additive exPlanation) 388
using, in pandas 187 simple model 324
using, in Python 186 simple RNN 156, 157
Region-based Convolutional Neural Network Sklearn 72
(R-CNN) 112 SpaCy
regular expressions 185 about 168-204
regularization Doc instance 168
about 92 Vocab class 168
in Keras 94, 95 Spearmint
L1 regularization 93 reference 318
L2 regularization 93 stationarity
reinforcement learning (RL) issues, avoiding 128
about 6, 278-290 significance 127
conversion, to supervised learning with types 126
Q-learning 281-283 Stochastic Gradient Descent (SGD) 35
effectiveness of data 7 stochastic volatility
frontiers 315 reference 415
machine learning models 8 structured data 41, 42
reward functions, designing supervised learning 4, 266
human preferences, learning 314 Synthetic Minority Over-sampling Technique
inverse reinforcement learning (IRL) 314 (SMOTE) 61
manual reward shaping 313 systematic error 366
robost RL, creating 315
RL engineering T
best practices 313
reward functions, designing 313 Tabotea project 214
rule-based matching TensorFlow 72
about 180, 181 tensors 37-39
combining, with learning based systems 185 Term Frequency, Inverse Document Frequency
custom functions, adding to (TF-IDF) 193
matchers 182, 183 test set 60
matchers, adding to pipeline 184 text classification task 188
[ 434 ]
Index
U
unsupervised learning 5, 266
V
vanishing gradient problem 347
variance stationarity 126
variational autoencoders (VAEs)
about 242
custom loss, creating 247, 248
Kullback-Leibler divergence 245-247
Lambda layer, using 244
MNIST example 243
used, for end-to-end fraud
detection 250, 251
using, for data generation 248, 250
using, in time series 251-253
[ 435 ]