Previsão de séries temporais multivariadas utilizando modelo híbrido interpretável com árvores de decisão fuzzy

Abstract

The present work aims to combine the idea of fuzzy systems with existing decision trees algorithms in order to design a multivariate forecasting method based on a rule set. In this way, a method which has explainability and interpretability based on the model’s input parameters is obtained, a property that is advantageous when compared to other various methods that provide mathematical expressions or black box algorithms (high complex algorithms), such as the neural networks models. The project was conceived in python language, and some other libraries and frameworks already designed in python, which allowed the creation of functions that can combine decision trees with fuzzyfied attributes. For instance, it was used the scikit-learn librarie, which has decision trees algorithms already implemented and the pyFTS library, which enables the transformation of numerical time series (output variable) into a fuzzy time series, and likewise transforms numerical and categorical input attributes into fuzzy time series. When performing this procedure, classification decision trees were used to generate “if-then” rules with fuzzy attributes both in their antecedent and in their consequent. After training a tree and obtaining the rules or graph from the tree itself as a knowledge base, defuzzification functions are used in order to transform fuzzy values into numerical ones and then predict the next time series value. In addition, the results of the method are competitive with other multivariate time series forecasting models, and it presents the flexibility to be applied in several series such as the following: electricity consumption, metro service demand, bovespa future contracts (B3), QoS (Quality of Service) and river regime. In addition, it was possible to verify that the production of rules as a knowledge base allowed an explanation for the movements of the series, as well as understanding which attributes are determinant to the movements of each series, and how these influence the series over time. Finally, it is worth noting that the obtained method has good results when considering the regression metrics used as MAE, RMSE, MAPE, SMAPE, among others. Beyond the good accuracy, as the model is composed of rules, it is possible to generate knowledge and understand how the values of the series are generated over time.