Supervised Classification with Gaussian Networks. Filter and Wrapper Approaches

Bayesian network based classifiers are only able to handle discrete variables. They assume that variables are sampled from a multinomial distribution and most real-world domains involves continuous variables. A common practice to deal with continuous variables is to discretize them, with a subsequent loss of information. The continuous classifiers presented in this paper are supported by the Gaussian network paradigm, which assumes that variables follow a Gaussian distribution. A great advantage of Gaussian network is that they need O(n) parameters to model a complete graph. This work shows how classifiers, supported by the Bayesian network paradigm, can be adapted to deal with continuous variables without discretizing them. In addition, two novel classifier learning algorithms are introduced. The presented learning algorithms are ordered and grouped according to their structural complexity: from the simplest naive Bayes structures to k-dependence Bayesian classifiers and semi naive Bayes. Moreover, for each structure a filter and wrapper approaches are presented. All these classifiers are empirically evaluated using the Brier score and the predictive accuracy. The obtained results with both scores suggest that semi naive Bayes is the best classifier.