Scribd Logo
Upload

Welcome to Scribd!

  • Interpretability With Diversified-By-Design Rules Skope-Rules A Python Package

    Uploaded by

    Hao OuYang
    97 views1 page
    Skope-Rules is a Python package that generates an interpretable rule-based classifier. It aims to learn decision rules for detecting instances of a target class with high precision. Skope-Rules differs from similar rule-learning models by applying a semantic deduplication approach to select rules based on their common variables rather than feature selection. It can be used for binary classification, describing clusters in clustering models, or understanding predictions from other complex classifiers. The package is hosted on Scikit-Learn-Contrib and its code and documentation are available on GitHub.

    Original Description:

    Original Title

    Interpretability with diversified-by-design rules skope-rules a python package

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document
    Skope-Rules is a Python package that generates an interpretable rule-based classifier. It aims to learn decision rules for detecting instances of a target class with high precision. Skope-Rules differs from similar rule-learning models by applying a semantic deduplication approach to select rules based on their common variables rather than feature selection. It can be used for binary classification, describing clusters in clustering models, or understanding predictions from other complex classifiers. The package is hosted on Scikit-Learn-Contrib and its code and documentation are available on GitHub.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    97 views1 page

    Interpretability With Diversified-By-Design Rules Skope-Rules A Python Package

    Uploaded by

    Hao OuYang
    Skope-Rules is a Python package that generates an interpretable rule-based classifier. It aims to learn decision rules for detecting instances of a target class with high precision. Skope-Rules differs from similar rule-learning models by applying a semantic deduplication approach to select rules based on their common variables rather than feature selection. It can be used for binary classification, describing clusters in clustering models, or understanding predictions from other complex classifiers. The package is hosted on Scikit-Learn-Contrib and its code and documentation are available on GitHub.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    of 1

    Interpretability with

    diversified-by-design rules;
    skope-rules, a python package
    [email protected], [email protected], [email protected]
    with the participation of n. goix, f. gardin, j-m. schertzer

    Motivation Skope-Rules: an interpretable rule-based classifier


    Complex Machine Learning models some- Skope-Rules, the proposed interpretable However, our approach mainly differs in the
    times achieve high performance, but lead model, aims at learning decision rules for way that decision rules are chosen: semantic
    to opaque decisions. Due to regulation or "scoping" a target class, i.e. detecting in- deduplication based on variables com-
    severe consequences on errors, more inter- stances of this class with high precision. posing each rule as opposed to L1-based
    pretability is often necessary [1]. Here, The problem of generating such rules has been feature selection (RuleFit).
    we present Skope-Rules, a rule-based inter- widely considered, see e.g. RuleFit [2], Slip-
    pretable model. per [3], LRI [4], MLRules [5].

    Methodology Implementation
    • Bagging estimator training: Multi- • Semantic deduplication: A similarity Skope-rules is a
    ple decision tree classifiers, and poten- filtering is applied to maintain enough Python package
    tially regressors (if a sample weight is diversity among the rules. The similar- hosted on Scikit-
    applied), are trained. Note that each ity measure of two rules is based on the Learn-Contrib
    node in this bagging estimator can be number of their common terms. A term under the 3-Clause
    seen as a rule. is a variable name combined with a com- BSD license.
    • Performance filtering: Out-of-bag parison operator (< or >).
    precision and recall thresholds are ap- Code and documentation are available here:
    plied to select best rules. https://github.com/
    scikit-learn-contrib/skope-rules
    Installation: pip install skope-rules

    Usage of Skope-Rules
    Skope-Rules is suited for different applica-
    tions and can be used:

    As a global interpretable model:


    This is the natural use of Skope-Rules, for a
    binary classification task. Rules isolate the 1s
    from 0s.
    Example
    • Line 4-5: Model is trained through As a cluster describer:
    standard scikit-learn API. In a clustering task, Skope-Rules is very use-
    • Line 8: Predictions are made and can ful to describe each segment. Each cluster can
    be used to evaluate performances of be post-processed and approximated with a
    rules combined. set of interpretable rules.
    • Line 11: The computed rules are stored
    with their out-of-bag standalone perfor- As a distribution queue describer:
    mances (see below). Skope-Rules is also relevant to describe how
    any subsample differs from a population.
    More precisely, it can be used to describe sub-
    samples defined by highest (or lowest) values
    of a given variable.
    In particular, Skope-Rules has revealed to be
    effective at understanding highest (or lowest)
    prediction scores of an other complex classi-
    fier.

    References Perspectives
    [1] Doshi-Velez et al. Accountability of AI Under the Law: The Role of Explanation, 2017 • Mathematical formalization
    [2] Friedman and Popescu. Predictive learning via rule ensembles,Technical Report, 2005
    • Improvement of the prediction API
    [3] Cohen and Singer. A simple, fast, and effective rule learner, National Conference on AI, 1999
    when combining rules
    [4] Weiss and Indurkhyar. Lightweight rule induction, ICML, 2000
    [5] Dembczynski, Kotlowski and Slowinski. Maximum Likelihood Rule Ensembles, ICML, 2008 • Develop paralleled implementations

    You might also like