The Learning and Optimizing of Markov Network Classifiers Based on Dependency Analysis

To decomposable probability model, it is proved that the Markov network classifier is optimal under zero-one loss. At present, the algorithms of learning the structure of Markov network classifier are inefficient and unreliable. In this paper, a new method of learning the structure of Markov network classifier is presented. The classifier structure is built by combining basic dependency relationship between variables, basic structure between nodes and the idea of dependency analysis. And Markov network classifier is optimized by removing unrelated and redundancy attribute variables to improve the ability of withstanding noise and predicting. A contrast experiment about the accuracy of classifiers is done by using artificial and real data. Experimental results show high classing accuracy of optimized Markov network classifier.