The stability of a restricted Bayesian network: an empirical investigation

Abstract

The stability is an important criterion of evaluating classification algorithms. Bayesian network classifier is one of the most popular classification methods, however, its stability is rarely studied. Tree augmented naive Bayes (TAN), and a restricted Bayesian network, have demonstrated stronger whole performance than the other Bayesian classification methods. The purpose of this paper is to study the stability of TAN. Bayesian network classification method and TAN model are firstly introduced, and then an empirical investigation comparing the stability of several typical classification approaches (decision tree, naive Bayes) with TAN are described. Experimental results show that tree augmented naive Bayes network classifier is stable.