Feature Selection Methods in the Framework of mRMR

Feature selection (FS) plays an important role in machine learning. FS under minimum redundancy maximum relevance framework based on mutual information behaved well according to existing researched. This paper focus on the validity of the Min-Redundancy Max-Relevance (mRMR) framework with some traditional correlative criteria, such as Spearman coefficient, distance correlation (dCor), and maximal information coefficient (MIC), etc. Experimental results show that mRMR can bring encouraging feature selection result compared with the traditional K-BEST feature selection method, no matter which criterion is adopted and the classification accuracy of these criteria is improved under the mRMR framework.