Feature selection using a mutual information based measure

In this paper, we discuss the problem of feature selection for the purpose of classification and propose a solution based on the concept of mutual information. In addition, we propose a new evaluation function to measure the ability of feature subsets in distinguishing between class labels. The proposed function is based on the information gain and takes into consideration how features work together. Finally, we discuss the performance of this function compared to that of other measures which evaluate features individually.