Feature selection based on complementarity of feature classification capability

Along with emergence of the high dimensionality of data, feature selection techniques are getting more significant to learning algorithms. Many metrics have been introduced in feature selection. Among them, mutual information is a highlighted one and has been developed during the past years. In this paper, a novel feature selection method based on the measurement of complementarity of feature classification capability is presented. By measuring the relevance between features and classes in terms of normalized mutual information, and maximizing complementarity of classification capability of features, the proposed method can sift relevant features and avoid irrelevant and redundant features simultaneously. The proposed method is compared with the related studies through applied to three different classifiers on five VCI datasets and six gene expression datasets. The experimental results showed that the proposed method achieved a better performance while involving a smaller number of features in most cases.