Attribute weighted methods for Bayesian Network Classifiers

Bayesian Network (BN) is a popular classification technique for data mining and machine learning, which is based on the attribute independence assumption. Researchers have proposed out many effective methods to improve the performance of Bayesian Network Classifiers (BNC) by lowering its primary weakness—the assumption that attributes are independent given the class, such as backwards sequential elimination method, lazy elimination method and so on. Recently, some new technology has been used to improve the accuracy of BNC. In this paper, our research is focused on weighted method in Attribute Weighted Bayesian Network Classifiers (WBNC). We firstly make a detailed experimental analysis to investigate the effect of attribute weight on the various BNC, such as NB, AODE and HNB. Then we present an improved weighted attribute method in connection with NBC based on differential evolution algorithms, simply DE-WBNC. In DE-WBNC, we use differential evolution algorithms to determine the weights of attributes and then use these weights in our previously developed WBNC. We evaluate the performance of new configuration DE-WBNC on the whole 36 standard UCI data sets in Weka system. Experimental results confirm the effectiveness of the new methods. Comparisons with state-of-the-art attribute methods via the based BNCs (NB, AODE and HNB), DE-WBNC, are provided, highlighting advantages on classification accuracy of the methods proposed.