Um algoritmo evolutivo para indução de árvores de regressão robusto a valores ausentes

Abstract

A common concern in many fields of knowledge involves problems of low quality data, such as noise and missing data. In the machine learning area, for example, missing data has generated serious problems in the knowledge extraction process, hiding important information about the dataset, skewing results and affecting the accuracy of the induced models. In order to deal with these problems, much has been discussed in the literature about missing values treatment strategies, either by preprocessing tasks or by the implementation of robust algorithms to missing data. In this thesis, we introduce a new evolutionary algorithm for induction of regression trees, including multiple strategies in its evolutionary cycle for dealing with missing data. Aiming to make a comparative analysis, we evaluated six traditional regression algorithms over 10 public datasets artificially modified to present different levels of missing data. Results from the experimental analysis show that the proposed solution presents a good trade-off between model interpretability and predictive performance, especially for datasets with more than 40% of missing data.