A Comparative Study of Redundant Feature Detection based Feature Selection Methods

As a high dimensional problem, analysis of large-scale data sets is a challenging task, where many weakly relevant or redundant features hurt generalization performance of classification models. In order to solve this problem, many effective feature selection methods have proposed to eliminate redundant features in recent years. However, the comparative performances of these redundant feature detection based methods have not been reported yet, which makes the choice of feature selection method relatively difficult for many real applications. The paper presents a novel comparative study of redundant feature detection based feature selection methods. Experiments on several benchmark data sets demonstrate the comparative performances of some state-of-the-arts methods. Based on the extensive empirical results, the minimum Redundancy-Maximum Relevance (mRMR) method has been found to be the best one among all compared feature selection models.