Attribute relevance in multiclass data sets using the naive Bayes rule

Feature selection using the naive Bayes rule is presented for the case of multiclass data sets. In this paper, the EM algorithm is applied to each class projected over the features in order to obtain an estimation of the class probability density function. A matrix of weights per class and feature is then obtained, where it collects the level of relevance of each feature for the different classes. We show different ways to extract this information and compare the behavior of the ranking of relevance obtained applying the naive Bayes and K-NN classifiers.