Recursive Feature Selection Based on Minimum Redundancy Maximum Relevancy

Abstract

Minimum redundancy maximum relevancy (mRMR) is one of the successful criteria used by many feature selection techniques to evaluate the discriminating abilities of the features. We combined dynamic sample space with mRMR and proposed a new feature selection method. In each iteration, the weighted mRMR values are calculated on dynamic sample space consisting of the current unlabelled samples. The feature with the largest weighted mRMR value among those which can improve the classification performance is preferred to be selected. Five public data sets were used to demonstrate the superiority of our method.