Abstract
Feature selection is an important preprocessing step in pattern recognition. In this paper, we presented a new feature selection approach in two-class classification problems based on information theory, named minimum Distribution Similarity with Removed Redundancy (mDSRR). Different from the previous methods which use mutual information and greedy iteration with a loss function to rank the features, we rank features according to their distribution similarities in two classes measured by relative entropy, and then remove the high redundant features from the sorted feature subsets. Experimental results on datasets in varieties of fields with different classifiers highlight the value of mDSRR on selecting feature subsets, especially so for choosing small size feature subset. mDSRR is also proved to outperform other state-of-the-art methods in most cases. Besides, we observed that the mutual information may not be a good practice to select the initial feature in the methods with subsequent iterations.