关键词: 半朴素贝叶斯分类器, 互信息, 匹配

Abstract: Because the class-conditional independence assumption which is the mainly feature of Naive Bayesian classifier is a so strong assumption and difference appears between datasets, the class-conditional independence assumption becomes the entry point of improvement methods. But some researches indicate that the violations of independence assumption do not make so much influence to the classifier as expected. This paper proposes a conditional entropy matching half-naive Bayesian classifier for the purpose of lower posterior probability estimation error. Experiments show that this method can effectively improve the performance of naive Bayesian classifier.

Key words: semi-naive Bayesian classifier, mutual information, matching