A Smart Design for the TAN Classifier

The naive Bayesian classifier is widely used because of itpsilas simplicity and effectiveness. But it has a strict assumption about the independence for each attribute, which is not obviously hold in real world domains. Many efforts have been made to relax the independence and improve the performance of the naive Bayesian classifier. Tree Augmented Naive Bayes (TAN) classifier was proved to be one of the best methods. In this paper, we analyze the implementations of distribution-based TAN classifier and the classification-based TAN classifier. Then we utilize the information theory to compute the influence between two attributes, and then proposed a new heuristic searching measurement for the tree structure. The experimental results have shown the advantage of the new classifier.