Redes Bayesianas para classificação com aprendizado via scoring and restrict: método, aplicação e comparação com métodos tradicionais

Abstract

This work is an investigation towards the behavior of discrete Bayesian Networks (BN) which aims to solve classification problems. This methodology is based on graphs and probability theories, and it is defined to be a probabilistic graphical model that allows the relationship visualization among (random) variables and, in general, simplifies the understanding of complex domains. To understand their performance, some classifiers were selected to be compared, such as Naïve Bayes (NB), Tree Augmented Naïve Bayes (TAN), K-Dependence Bayesian Network (KDB), Bayesian Network Augmented Naïve Bayes (BAN), and Averaged One-Dependence Estimator (AODE). In general, the performance of the ensemble classifier AODE outperforms the others. In addition, a hybrid method for structure estimation is proposed and it aims the parsimonious classification. The simulation studies show the new method fits the expectation of increase the prediction performance also, balance the number of connections among variables and, the applications in real datasets support the better understanding of the new approach. Finally, a combination of the classifiers via stacking was presented and it indicated an increase in their performances when they were compared to the methods themselves.