Using Relative Classification Probability to Increase Accuracy of Restricted Structure Bayesian Network Classifiers

Abstract

Bayesian networks have been used widely in probabilistic representation and reasoning. Meanwhile, it has been shown that Bayesian classifiers are competitive with many state-of-the-art classifiers. In this paper we present an approach that provides a good tradeoff for the Bayesian network classifier between the number of classified instances and classification accuracy, based on a measure of relative classification probability (RCP). Experiments on benchmark datasets show good support for our hypothesis. The same classifier could reach much higher accuracy over a subset of the original dataset. For most datasets, classification accuracy of the same classifiers can rise high without excluding many instances. The empirical study shows that this idea works well especially for the multiclass classification case.