A Max-relevance-min-divergence Criterion for Data Discretization with Applications on Naive Bayes