Business Information Systems Workshops: BIS 2019 International Workshops, Seville, Spain, June 26-28, 2019, Revised Papers Witold Abramowicz
Business Information Systems Workshops: BIS 2019 International Workshops, Seville, Spain, June 26-28, 2019, Revised Papers Witold Abramowicz
https://textbookfull.com/product/biota-grow-2c-gather-2c-cook-
loucas/
https://textbookfull.com/product/business-information-
systems-22nd-international-conference-bis-2019-seville-spain-
june-26-28-2019-proceedings-part-ii-witold-abramowicz/
https://textbookfull.com/product/business-information-
systems-22nd-international-conference-bis-2019-seville-spain-
june-26-28-2019-proceedings-part-i-witold-abramowicz/
https://textbookfull.com/product/business-information-systems-
workshops-bis-2014-international-workshops-larnaca-cyprus-
may-22-23-2014-revised-papers-1st-edition-witold-abramowicz/
https://textbookfull.com/product/business-information-systems-
workshops-bis-2016-international-workshops-leipzig-germany-
july-6-8-2016-revised-papers-1st-edition-witold-abramowicz/
https://textbookfull.com/product/euro-par-2019-parallel-
processing-workshops-euro-par-2019-international-workshops-
gottingen-germany-august-26-30-2019-revised-selected-papers-
ulrich-schwardmann/
https://textbookfull.com/product/business-information-
systems-23rd-international-conference-bis-2020-colorado-springs-
co-usa-june-8-10-2020-proceedings-witold-abramowicz/
Witold Abramowicz
Rafael Corchuelo (Eds.)
Business
Information Systems
LNBIP 373
Workshops
BIS 2019 International Workshops
Seville, Spain, June 26–28, 2019
Revised Papers
123
Lecture Notes
in Business Information Processing 373
Series Editors
Wil van der Aalst
RWTH Aachen University, Aachen, Germany
John Mylopoulos
University of Trento, Trento, Italy
Michael Rosemann
Queensland University of Technology, Brisbane, QLD, Australia
Michael J. Shaw
University of Illinois, Urbana-Champaign, IL, USA
Clemens Szyperski
Microsoft Research, Redmond, WA, USA
More information about this series at http://www.springer.com/series/7911
Witold Abramowicz Rafael Corchuelo (Eds.)
•
Business
Information Systems
Workshops
BIS 2019 International Workshops
Seville, Spain, June 26–28, 2019
Revised Papers
123
Editors
Witold Abramowicz Rafael Corchuelo
Poznań University of Economics University of Seville
and Business Seville, Spain
Poznan, Poland
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
In 2019 we had a great opportunity to organize the 22nd edition of the International
Conference on Business Information Systems (BIS 2019), that has grown to be a
well-renowned event for scientific and business communities. This year the main topic
of the conference was “Data Science for Business Information Systems.” The con-
ference was jointly organized by the University of Seville, Spain, and the Poznań
University of Economics and Business, Poland, and was held in Seville, Spain.
During each edition of the BIS conference series we make the effort to provide an
opportunity for discussion about up-to-date topics from the area of information systems
research. However, there are many topics that deserve particular attention. Thus, a
number of workshops and accompanying events are co-located with the BIS conference
series. The workshops give researchers the possibility to share preliminary ideas and
initial experimental results, and to discuss research hypotheses from a specific area of
interest.
Nine workshops and one accompanying event took place during BIS 2019. We were
pleased to host well-known workshops such as AKTB (11th edition), BITA (10th
edition), iCRM (4th edition), and iDEATE (4th edition), as well as relatively new
initiatives such as ISAMD, DigEx, BSCT, SciBOWater, and QOD. Each workshop
focused on a different topic: knowledge-based business information systems (AKTB),
challenges and current state of business and IT alignment (BITA), integrated social
CRM (iCRM), Big Data and business analytics ecosystems (iDEATE), Blockchain
(BSCT), digital customer experience (DigEx), maritime systems (ISAMD), water
management (SciBOWater), and data quality (QOD).
Additionally, BIS 2019 hosted a Doctoral Consortium. It was organized in a
workshop format, thus the best papers from this event are included in this book.
Moreover, all authors had the possibility to discuss their ideas on PhD thesis and
research work with a designated mentor.
The workshop authors had the chance to present their results and ideas in front of a
well-focused audience; thus the discussion gave the authors new perspectives and
directions for further research. Based on the feedback received, authors had the
opportunity to update their workshop articles for the current publication. This volume
contains 57 articles that are extended versions of papers accepted for BIS workshops.
In total, there were 139 submissions for all mentioned events. Based on the reviews, the
respective workshop chairs accepted 57 in total, yielding an acceptance rate of 41%.
We would like to express our thanks to everyone who made BIS 2019 workshops
successful. First of all, our workshops chairs, members of the workshop Program
vi Preface
AKTB Workshop
BITA Workshop
BSCT Workshop
DigEX Workshop
iCRM Workshop
iDEATE Workshop
ISAMD Workshop
QOD Workshop
SciBOWater Workshop
Doctoral Consortium
Virgilijus Sakalauskas
Dalia Kriksciuniene
Organization
Chairs
Program Committee
Tsung-Nan Chou(&)
Abstract. Machine learning and deep learning are all part of artificial intelli-
gence and have a great impact on marketing and consumers around the world.
However, the deep learning algorithms developed from the neural network are
normally regarded as a black box because their network structure and weights
are unable to be interpreted by a human user. In general, customers in the
banking industry have the rights to know why their applications have been
rejected by the decisions made by black box algorithms. In this paper, a practical
grafting method was proposed to combine the global and the local models into a
hybrid model for explainable AI. Two decision tree-based models were used as
the global models because their highly explainable ability could work as a
skeleton or blueprint for the hybrid model. Another two models including the
deep neural network and the k-nearest neighbor model were employed as the
local models to improve accuracy and interpretability respectively. A financial
distress prediction system was implemented to evaluate the performance of the
hybrid model and the effectiveness of the proposed grafting method. The
experiment results suggested the hybrid model based on the terminal node
grafting might increase the accuracy and interpretability depending on the
chosen local models.
1 Introduction
The machine learning and artificial intelligence (AI) technology have been successfully
applied to many financial services to solve various decision-making problems. How-
ever, the pervasive deep learning techniques developed from the neural network is
considered as a black box and cannot be explained by humans because their network
architecture and weights are difficult to be interpreted. Consequently, their processed
results are unable to confirm whether the decision-making and analysis are reasonable.
In addition, in the case of a consumer loan application, if the financial institution rejects
an applicant’s loan, the borrower has the rights to ask for the reason why the loan is not
approved. In such a situation, the credit decision generated by the automated decision
system using the deep learning algorithm needs to explain how the judgment process
work. In order to avoid possible irreparable harm from the automated decision-making
systems, the General Data Protection Regulation (GDPR) approved by the European
Union in 2016 and enforced in 2018 stipulates that rejected applicants have the rights to
know the details of the process leading to their rejection [1]. Therefore, for the most
common applications of machine learning in finance, legal consultation, and medical
diagnosis, the user must be informed with the reasons for decision or judgment.
Otherwise, even the decision is correct, or the prediction model is highly accurate, the
result is still doubtful.
Unlike the difficulty of understanding why artificial intelligence has failed, humans
can understand and discover their decision and judgment errors and then correct them
to avoid making the same mistakes. Therefore, many researchers have tried to develop
different methods to make the black boxes of artificial intelligence understandable
without affecting their efficiency and performance. As a result, the research area such as
the explainable AI (XAI), which is concerned with how to make humans understand
the reasons for judgment under artificial intelligence, is getting more attention. How-
ever, while using XAI models for prediction, these models need to deal with the
tradeoff between model interpretability and accuracy [2]. Generally, an effective XAI
model could be used to explain the decision rules generated, but the performance of the
model might be affected. Since the artificial intelligence is originally designed to solve
a large amount of data analysis problems that humans are difficult to handle, whether
the AI models can be understood is closely related to the complexity of the algorithms
they used. In general, the more complex a model is, the more difficult to interpret its
decision-making and reasoning processes. How to make artificial intelligence inter-
pretable without affecting its performance for an automated decision-making system
will be a challenge for researchers. In deep learning techniques, the generative
adversarial network (GAN) could be used to train two neural networks against each
other to maximize learning outcomes and explore the internal workings of neural
networks to provide interpretability [3]. However, the training of GAN is time-
consuming and might be unstable. In contrast, the decision tree generates a set of
decision rules that are easy for the human to interpret, and the algorithm could be one
of the solutions that help the banking industry to be compliant with the GDPR
requirements. Although the concepts of explainability and interpretability are different
for XAI applications, both are used interchangeably in this paper because the applied
decision trees are interpretable models that also provide understandability to know how
the results are achieved. The following sections introduce the methods and strategies
used in this study briefly.
Some machine learning algorithms such as decision trees, Bayesian inference or Sparse
linear models (SLIM) [4] intrinsically provide human interpretability that is easy to
understand. Alternatively, many research approaches attempt to realize how the
machine learning model work by examining their global or local interpretability, and
some others use model-specific or model-agnostic to distinguish the way in which
interpretability can be generated. In general, model-agnostic approaches such as partial
dependence plot (PDP), Shapley value and local interpretable model-agnostic expla-
nations (LIME) provide interpretability that is compatible with any classification,
A Practical Grafting Model Based Explainable AI 7
regression, or reinforcement learning models [5]. The PDP model [6] generally changes
the value of a certain characteristic variable one by one while controlling other vari-
ables, and interprets the relationship between the characteristic variable and the target
variable by a line graph. Comparatively, Lundberg and Lee [7] proposed the Shapely
explanations method, using the Shapley value designed by Lloyd Shapley [8] to
explain the importance of feature variables. In general, the deep learning models use the
training sample to learn the relationship between the input and output variables. If the
sample data increases, the mapping of relationship becomes more complicated and
makes the constructed model difficult to understand. The LIME algorithm uses a small
number of regional samples to construct a simple local model, and the importance of
the variable with respect to a testing sample is calculated by the local model rather than
the global model previously trained. In brief, this algorithm uses the local model as a
proxy model to interpret the original black-box global model [9]. Since the LIME
algorithm is model-agnostic, the global model can be any machine learning or deep
learning algorithm, while the local model can apply a simple linear regression or a
decision tree to fit the target values predicted by the original global model.
This paper proposed a practical grafting method with intent to overcome the
tradeoff between accuracy and interpretability in machine learning based on the inte-
gration of the global model and the local model. In the first place, each of five common
machine-learning models was applied as a global model to evaluate its individual
performance. Especially, the decision tree (DT) and fuzzy decision tree (FDT) were
used as the global modes due to their explainable ability. Both models worked as a
skeleton or blueprint to concatenate the local models. To balance the tradeoff between
accuracy and interpretability, the deep neural network (DNN) and k-nearest neighbor
methods (KNN) were chosen as local models, where the former intended to increase
the accuracy and the latter attempted to increase the interpretability. Finally, the ter-
minal node grafting method that integrated the global and the local models based on
three fusion strategies were implemented to examine the performance of the hybrid
model. The methods and strategies used in this study were briefly described as
following.
but also as a local model in conjunction with the DT or the FDT global model to
construct a hybrid system. The number of hidden layers and the corresponding pro-
cessing units within the sequential DNN model [13] were examined with different
experimental configurations. Furthermore, the dropout layers with a cut-off value of 0.1
were used to control the overfitting problem. The activation functions such as Sigmoid,
Tanh and ReLU were assessed in different hidden layers and another SoftMax function
was employed in the last fully connected layer.
As shown in Fig. 1, for any decision path ended with a terminal node, the subset of
training data partitioned by this path was collected to train another specified local
model if it performed poor accuracy. The model interpretability remained unchanged if
using the inexplicable DNN model as a local model, while using the KNN model might
increase the model interpretability. The other grafting method was the internal node
grafting illustrated in Fig. 2. The internal nodes (or decision nodes) of a decision tree
performing lower accuracy were removed and replaced with a local model such as the
DNN or the KNN model. The branches of a decision tree resembled the clustering
function to partition the training data into smaller subsets for training the local models.
MF1, MF2, and MF3. The decision rules generated by the DT and the FDT models
comprised 6 and 10 rules, respectively. As the excerpted rules illustrated in Fig. 3, the
fuzzy rules derived from the FDT model were controlled by the maximum of four input
variables in the non-crisp antecedents to reduce the complexity of the global model.
Generally, the DT model searched for a single suitable if-then rule from its trained
rule set and applied it to the prediction of a testing sample. In contrast, the FDT model
required to aggregate multiple fuzzy rules to evaluate the overall effect on that pre-
diction. Therefore, this difference raised two challenges for implementing the FDT
model as a global model in grafting. The first challenge concerned about how to use the
multiple rules of the FDT model to train each of their corresponding local models. In
terminal node grafting, the training data that complied with a single decision rule of the
DT model was directly partitioned as an individual subset for training the local model.
In contrast, implementing the FDT model as a global model required to combine
several subsets of training data for the local models according to the aggregation of
multiple fuzzy rules. The other challenge was how to integrate the global model with
the local model to achieve the best synthesis results. If the FDT model was used as a
global model, each of its activated fuzzy rules imposed a different weighting effect on
the prediction of the testing sample.
Accordingly, three fusion methods were developed to quantify the effect of fuzzy
rules on their concatenated local models and calculate the synthetic outcome for the
testing data. The first fusion method was a weighted sum approach. The evaluation of a
testing sample was calculated by the weighted sum of the outputs from its corre-
sponding local models, where the weights were the relative importance derived from
the activated fuzzy rule associated with the local model. In addition, the second fusion
method evaluated the testing sample based on the average output of the local models
triggered by their corresponding activated fuzzy rules from the global model. More-
over, because each of the activated fuzzy rules represented a weighted vote to its
associated local model, the evaluation of a testing sample in the third fusion method
was decided by the activated fuzzy rule that acquired the relative majority. In this
paper, the FDT model implemented with the KNN model based on three fusion
methods were denoted as the FDT-KNN (1) to FDT-KNN (3) models. Comparatively,
A Practical Grafting Model Based Explainable AI 11
the FDT model integrated with the DNN model were denoted as the FDT-DNN (1) to
FDT-DNN (3) models.
3 Experiment Results
To evaluate the integrated performance of the global and the local modes based on
different grafting and fusion methods, a financial distress prediction system was
implemented with various experimental configurations. The dataset for experiments
comprised the financial and corporate information of 32 distress and another 32 healthy
companies declared by the authoritative Taiwan Economic Journal (TEJ) in accordance
with various criteria and definitions during the period of 2008 and 2015. Furthermore,
the dataset consisted of 15 independent variables were subdivided into 80% for the
training dataset and another 20% for the testing dataset. The DT, FDT and KNN
models were all regarded as interpretable models regardless of whether they were used
as the global or local model. Totally, five experiments were established to evaluate the
performance of the model grafting strategies based on four metrics including accuracy,
kappa, sensitivity, and specificity. Considering the tradeoff between the accuracy and
interpretability for training tree-based global models, the complexity parameter
(CP) controlling the number of splits in a decision tree by examining the misclassifi-
cation error for each branch was evaluated with 0.01 and 0.001, and both were denoted
as the DT1 and DT2 models in the first experiment. Moreover, the minimum number of
samples in a terminal node and the maximum depth and height of a tree were also
controlled to construct a concise global model. As shown in Table 1, the experiment
result reported that the DT2 model demonstrated better training accuracy of 0.89, only
5% difference from the simpler DT1 model. However, the DT2 model created a larger
rule set of 12 if compared to the smaller DT1 model generating 6 rules only. Although
the accuracy of DT1 model was slightly lower than the DT2 model, the DT1 model was
selected as the global model to concatenate local models because it was easier to
interpret, and the local models might recover the minor decrease of accuracy during
grafting experiments.
The second experiment intended to compare the performance of the five common
models described in Sect. 2 prior to working as either a local model or global model.
As indicated in Table 2, the FDT model was inferior to the other four models although
all models achieved the accuracy over 0.70. The accuracy of the FDT model was 0.73,
a 6% deteriorated from the original DT model being reported. The reason why the DT
model performed better than the FDT model was that the latter was pruned to preserve
a similar tree structure as its counterpart DT model for further comparison of predictive
12 T.-N. Chou
accuracy after grafting, and the crisp inputs needed to be transformed into fuzzy inputs
for a fuzzy inference rule. As the random forest was an ensemble approach that
combined multiple decision trees into a single predictive model to reduce variance and
bias, the experiment result unquestionably suggested the RF model was the best model
and achieved an accuracy of 0.85, which outperformed all other common models.
The model grafting created a hybrid model that combined the global model with the
local model to increase the overall interpretability and accuracy. Both the DT and the
FDT models could be employed as a global mode because they provided better
interpretability for the human to understand. On the other hand, the KNN model used
as a local model with an attempt to increase more interpretability, while the DNN
model was applied to improve predictive accuracy. The performance of the terminal
node grafting based on using the DT model as a global model was reported in Table 3.
As opposed to the single DT model obtained the accuracy of 0.79, the DT-based
grafting result indicated a moderate improvement, and the accuracy was increased to
0.87 for concatenated with the DNN model and 0.82 for the KNN model respectively.
Although the DT-DNN model achieved higher accuracy, however, the DT-KNN model
gained more interpretability by slightly losing the accuracy of 5%.
Since the FDT model required to concatenate several local models together due to
the multiple fuzzy rules been activated simultaneously, three fusion methods were
applied to merge the outcomes of the FDT model with its corresponding local models.
By referring to Table 4, the accuracy of the FDT-DNN (3) model using the third fusion
method performed the lowest accuracy of 0.73, the same as the result of using the FDT
model alone. On the other hand, the FDT-DNN (1) model using the first fusion method
improved the accuracy to 0.88 and it was considered as the best model if compared to
the other two fusion methods.
A Practical Grafting Model Based Explainable AI 13
In the final experiment, the KNN model was employed as a local model in con-
junction with the FDT model to evaluate three fusion methods. As reported in Table 5,
the FDT-KNN (1) model outperformed the other models and achieved an accuracy of
0.84. However, this result was slightly lower than that of the FDT-DNN (1) model
using the same fusion method and suggested that the KNN model might not be able to
improve the predictive accuracy substantially. Although the DNN model aided the
hybrid model to improve the predictive accuracy, the KNN model implicated more
explanations could be obtained to learn how the results were achieved.
As a highlight, while using only a single model in the experiment, the RF model
achieved the highest predictive accuracy among five testing models. However, the
interpretability of the RF model was limited since the model applied an ensemble of a
number of decision trees. Moreover, by using a hybrid model that integrated a global
model with several local models, the predictive results varied and depended on which
fusion method was used in model grafting. As the experiment results suggested, the
first fusion method surpassed the other two fusion methods, and the combination of the
FDT and DNN models might outperform the single RF model in predictive accuracy.
Conversely, by using the FDT model in conjunction with the KNN model, the pre-
dictive accuracy was slightly worse than the RF model. If considering the DT model
instead of the FDT model to concatenate with the local models, the combination of the
DT and the DNN model also performed better than the single RF model. Consequently,
both the DT and FDT model grafted with the DNN model might be the favorable
hybrid solutions to increase predictive accuracy, even by losing a bit of interpretability
if compared with the KNN model.
4 Conclusion
global model and the local model together with the intent to improve the overall
interpretability and accuracy of the resulting hybrid model. Both the DT and the FDT
models were chosen as the global models due to their highly explainable ability. On the
other hand, the inexplicable DNN model and the explainable KNN model were applied
as local models to improve the accuracy and interpretability respectively. By
employing a signal model in the experiment, the result indicated the RF model was the
best model with the predictive accuracy of 0.85. However, if we implemented a hybrid
model based on three fusion methods in the terminal node grafting, the FDT-DNN
(1) model was considered as the most competent model under all experiments and
achieved the highest accuracy of 0.88. In addition, the experiment result also suggested
that both the DT and the FDT models were capable of providing the explainable ability
and could be combined with the DNN model to improve model performance by using
the proposed grafting technique. Although the KNN model might deliver additional
interpretability to the hybrid model, the predictive accuracy was slightly decreased.
This research work can be further extended to implement the internal node grafting and
compare its performance with that of terminal node grafting described in this study.
References
1. Goodman, B., Flaxman, S.: European union regulations on algorithmic decision-making and
a “right to explanation”. AI Mag. 38, 50–57 (2017)
2. Adadi, A., Berrada, M.: Peeking inside the black-box: a survey on explainable artificial
intelligence (XAI). IEEE Access 6, 52138–52160 (2018)
3. Karras, T., Aila, T., Laine, S., Lehtinen, J.: Progressive growing of GANs for improved
quality, stability, and variation. arXiv:1710.10196v3 [cs.ne] (2018)
4. Ustun, B., Rudin, C.: Supersparse linear integer models for optimized medical scoring
systems. Mach. Learn. 102(3), 349–391 (2016)
5. Došilović, F.K., Brčić, M., Hlupić, N.: Explainable artificial intelligence: a survey. In: 2018
41st International Convention on Information and Communication Technology, Electronics
and Microelectronics (MIPRO), pp. 0210–0215 (2018)
6. Berk, R., Bleich, J.: Statistical procedures for forecasting criminal behavior: a comparative
assessment. Criminol. Public Policy 12(3), 513–544 (2013)
7. Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. In:
Advances in Neural Information Processing Systems (NIPS 2017), vol. 30, pp. 4765–4774
(2017)
8. Roth, A.E. (ed.): The Shapley Value, Essays in Honor of Lloyd S. Shapley. Cambridge
University Press, Cambridge (1988)
9. Ribeiro, M.T., Singh, S., Guestrin, C.: Why should I trust you? Explaining the predictions of
any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pp. 1135–1144 (2016)
10. Rao, A., Golbin, I.: What it means to open AI’s black box. Next in Tech, PwC Blogs. http://
usblogs.pwc.com/emerging-technology/to-open-ai-black-box (2018), last accessed 2019/
04/14
11. Levashenko, V., Zaitseva, E.: Fuzzy decision trees in medical decision making support
system. In: Proceedings of 2012 Federated Conference on Computer Science and
Information Systems, pp. 213–219 (2012)
A Practical Grafting Model Based Explainable AI 15
12. Guillaume, S., Charnomordic, B.: Learning interpretable fuzzy inference systems with
Fispro. Int. J. Inform. Sci. 181(20), 4409–4427 (2011)
13. Liu, W., Wang, Z., Liu, X., Zeng, N., Liu, Y., Fuad, E.A.: A survey of deep neural network
architectures and their applications. Neurocomputing 234, 11–26 (2017)
14. Webb, G.I.: Decision tree grafting from the all tests-but-one partition. In: Proceedings of the
16th International Joint Conference on Artificial Intelligence (IJCAI), vol. 2, pp. 702–707
(1999)
Data Analytics in the Electronic Games
Abstract. This paper aims at the use of data analytics methods in mobile
games. The main goal was to predict future purchases of players in the selected
mobile game. The result presents the information about whether the player is
going to buy any of the offered bonus packages or not. This information is
crucial for marketing and possible ways of monetization. From the perspective
of data analytics, the goal is the creation of a classification model in line with the
CRISP-DM methodology. We used the following algorithms in the modeling
phase: Random forest, Naive Bayes, Linear regression, XGBoost, and Gradient
Boosting. All generated models were evaluated by contingency tables, which
presented models accuracy as the ration between successfully predicted values
to all predicted samples. The results are plausible and have the potential to be
deployed into practice as a baseline model or support for personalized marketing
activities.
1 Introduction
Everybody knows what mobile games are. We play them in our free time, on the bus,
some of us even at work. Mobile games sector is a highly competitive, and the
development of a successful game is a challenging and complex issue. A few years ago,
you came to the shop and bought a physical copy of the game. Nowadays, the process
is even more straightforward with the availability of online mobile stores. But the
players prefer another model for mobile games; it is called a fremium. The 2018 report
by Nielsen gaming research company SuperData has revealed that so-called freemium
games – games that are free to play but typically enable microtransactions for users to
purchase upgrades and gizmos – are dominating the entertainment market, generating a
massive $88 billion in the past year. This fact motivated us to start a collaboration with
one of the game development companies creating free-to-play games. During the
analytical process, we communicated intensively with the experts; not only during the
data understanding but also in the evaluation phase.
The paper is organized as follows: an introduction with motivation, related work,
methodology, and used methods. The description of the performed analytical project is
in line with the CRISP-DM methodology. The conclusion summarizes the results.
1
https://sensortower.com/blog/top-apps-games-publishers-2018.
Another random document with
no related content on Scribd:
Nestes dias festivaes
Com summa gala e grandeza,
Assistiu toda a nobreza
Dos homens mais principaes,
Ministros, officiaes
De guerra e damas mui bellas,
Que em palanques e janellas
Mostravam com arrebol,
Que estando alli posto o sol,[5]
Bem podiam ser estrellas.
Empinou-se-lhes a Ruça,
E de quatro companheiros,
Sem mais outros cavalleiros,
Fizeram a escaramuça:
O General se debruça,
Para mette-los bem nella,
Na janella com cautela;
Porém usou de revoltas,
Porque mettendo-os nas voltas,
Mandou fechar a janella.
A escaramuça acabada
Fizeram a cortezia,
E todo o povo seria
Vendo a janella fechada:
Nas voltas não viram nada;
Que com notavel trabalho
No ay hombre cuerdo á cavallo;
Porém depois que acabaram,
E o General não acharam,
Ficaram de vinha d’alho.
Com rostos descoloridos,
Desesperados agora
Iam por dentro e por fóra
Da propria côr dos vestidos:
Os que são desvanecidos
E, sem prudencia e razão,
Presumem mais do que são,
Emendem seus pensamentos,
Que para seus desalentos
É vivo o senhor d. João.
Seguia-se um cavalleiro
Ao famoso André Cavallo,
Que levou sem intervallo
De cada golpe um carneiro:
Tambem foi aventureiro
De um premio, mas com defeito
Dava ao corpo um grande geito,
E ficou passado e absorto
De que fosse ao premio torto
E o premio a outro direito.
Ao famoso Braz Rabello
Razão é de mestre o apode,
Pois dar dias sanctos póde
Nesta arte ao que fôr mais bello;
E si com louco desvelo,
Do que digo algum se abraza,
Attenda á razão que é raza,
E verá se faz espantos
Que dar possa os dias sanctos
Quem tem Domingas de Casa.
DECIMA
GLOZA
Já da primavera entrou
A alegre serenidade,
Com que toda a tempestade
Do triste inverno acabou:
Já Saturno declinou
Nas operações malignas:
Com influencias benignas
Jupiter predominante
Nos promette anno abundante
De flores e pedras finas.
Na terra já se experimentam
Virações tão temperadas,
Que as aves exterminadas
Tornar aos ninhos intentam:
Já não sentem, não lamentam
Tempestuosas ruinas;
Pois com salvas matutinas
Se mostram tão prasenteiras,
Que mais parecem caseiras
As aves que peregrinas.
Sua peregrinação
Influxo foi de Saturno,
Planeta sempre nocturno,
E muito importuno então:
Todas nessa conjuncção
Si os doces ninhos deixaram,
E tanto se receiaram
Do nocivo temporal,
Que escolhendo o menor mal
Aos montes se retiraram.
Porém tanto que sentiram
Haver no tempo mudança
Sem receio e sem tardança
Aos ninhos se reduziram:
Outros ares advertiram,
Outra clemencia notaram,
Com que alegres publicaram
Dos astros os movimentos,
E com festivos accentos
Nesta manhã já cantaram.