Scalable structure learning of K-dependence bayesian network classifier

Abstract

K-dependence Bayesian network classifier(KDB) has been widely used in data mining and machine learning. To enhance the expression ability and classification performance of KDB, the present study proposes a scalable k-dependence Bayesian classifier (SKDB), which is an extension of the classic KDB algorithm. SKDB introduces a novel attribute sorting method to pre-determine the optimal attribute order and a filtering mechanism to eliminate weak conditional dependence. Experimental results on 30 datasets from the University of California at Irvine (UCI) machine learning repository demonstrate that the above two techniques exert a positive impact on the classification performance of KDB. The proposed algorithm(SKDB) achieves better classification performance compared with several state-of-the-art BNCs (such as Naïve Bayes, tree-augmented Naïve Bayes, averaged one-dependence estimators, and KDB) in terms of 0-1 loss, bias, and variance.