Variable selection for Naïve Bayes classification

The Naïve Bayes has proven to be a tractable and efficient method for classification in multivariate analysis. However, features are usually correlated, fact that violates the Bayes’ assumption of conditional independence, may deteriorate method’s performance. Moreover, datasets often characterized by large number features, which complicate interpretation results as well slow down execution. In this paper we propose sparse version classifier is three properties. First, sparsity achieved taking into account correlation structure covariates. Second, different performance measures can used guide selection features. Third, constraints on groups higher interest included. Our proposal leads smart search, yields competitive running times, whereas flexibility terms measure integrated. findings show that, when compared against well-referenced feature approaches, proposed obtains regarding accuracy, times balanced datasets. case with unbalanced (or importance) classes, better compromise between rates classes achieved.