A differential evolution based feature selection approach using an improved filter criterion

Emrah Hancer, Bing Xue, Mengjie Zhang

Abstract: In filter feature selection, mutual information approaches have recently gained a high popularity among researchers. In these approaches, mutual information is commonly used to measure two components: the mutual relevance between each feature and the class labels, and the mutual redundancy between each pair of features. Despite their popularity, it has been pointed in the literature that such feature selection approaches may not fairly estimate the redundancy in high dimensional problems. To alleviate this problem, this paper proposes a new criterion, which uses the concepts of ReliefF instead of the mutual redundancy. Using the proposed criterion, a new differential evolution based filter feature selection approach is developed. The performance comparisons and analysis are conducted by comparing it with the most well-known mutual information feature selection (MIFS) criterion based on maximum-relevance and minimum-redundancy on the differential evolution framework. The results show that performing feature selection using the proposed criterion can generally achieve better classification performance and smaller feature subset size.