Fisher Score Based Naive Bayesian Classifier

Abstract

The nalve Bayesian classifier (NBC) is a simple yet very efficient classification technique in machine learning. But the unpractical condition independence assumption of NBC greatly degrades its performance. There are two primary ways to improve NBC's performance. One is to relax the condition independence assumption in NBC. This method improves NBC's accuracy by searching additional condition dependencies among attributes of the samples in a scope. It usually involves in very complex search algorithms. Another is to change the representation of the samples by creating new attributes from the original attributes, and construct NBC from these new attributes while keeping the condition independence assumption. Key problem of this method is to guarantee strong condition independencies among the new attributes. In the paper, a new means of making attribute set, which maps the original attributes to new attributes according to the information geometry and Fisher score, is presented, and then the FS-NBC on the new attributes is constructed. The condition dependence relation among the new attributes theoretically is discussed. We prove that these new attributes are condition independent of each other under certain conditions. The experimental results show that our method improves performance of NBC excellently.