A Clustering Based Bayesian Network Classifier

Building high-accuracy and efficient Bayesian network classifiers is a hot theme of Bayesian network classifier in recent years. It is often the case that building unrestricted Bayesian network classifier with large number of attributes is time-consuming and always gets poor result, since the searching space of network structure is huge. This paper proposes a clustering based Bayesian network structure learning algorithm(CBNA), which uses mutual information to measure the distances between attributes so as to divide them into groups by hierarchical clustering. Then the network searching is running under these low dimensional spaces while the relation of attributes of different group is ignored instead of finding the network from one high dimensional space. Experimental results suggested that this algorithm is more accurate and efficient when compared to other Bayesian network classifiers and enables to obtain the optimal structure by unrestricted searching.