Comparação entre redes bayesianas do tipo naive bayes baseadas na distribuição exponencial

Abstract

This paper compares and evaluates the precision of four evaluators of Naive Bayes based on the Exponentional Distribution. The acquired results were analysed according to the percentage of correct classification, the coefficient Kappa, its variance and the time of CPU consumed for each method. To analyse the performance of the methods, two groups of samples called training and testing samples, each having data from 1 to 4 dimensions with four different classes of performance. For the training samples with 10.000 observations were created for each and for the test samples were generated 16 samples with 30.000 observations. As result, the Method of Assessment and Training in the Fuzzy Exponentional Naive Bayes Network was significally better than the others. In relation to the time of CPU consumed, MATRENB managed to evaluate data that followed the Exponential Distribution, with four dimensions, therefore is a low computational cost application. It is worth to mention that from three or more dimensions most of the evaluators managed to acquire a percentual hits, with a moderate degree of substancial agreement, according to the Kappa coefficient.