Using optimal variables for Bayesian network classifiers

Using graphical models to represent independence structure in multivariate probability model has been studied since a few years. In this framework, Bayesian networks have been proposed as an interesting approach for uncertain reasoning. Within the framework of pattern recognition, many methods of classification were developed based on statistical data analysis. Belief networks were not considered as classifiers until the discovery that Naive Bayes, a very simple kind of Bayesian network, is surprisingly effective. In this paper, we propose to use belief networks classifiers with optimal variables that is to say networks which have to manage discrete and continuous variables.