Augmented Naive Bayesian Classifiers for Mixed-Mode Data

Conventional Bayesian networks often require discretization of continuous variables prior to learning. It is important to investigate Bayesian networks allowing mixed-mode data, in order to better represent data distributions as well as to avoid the overfitting problem. However, this attempt imposes potential restrictions to a network construction algorithm, since certain dependency has not been well modeled statistically. This work first introduces parametrical representation for dependencies among mixed-mode variables. It then proposes two associated structure learning algorithms, both intend to augment a naive Bayesian network. Experiments on a medical diagnosis application show that a naive Bayes with parametrical representation works significantly better than the one with pre-discretization, while a clique-augmented naive Bayes makes a slight further improvement.