Hybrid multi-objective PSO for filter-based feature selection

Abstract

This paper proposes a novel filter-based multi-objective particle swarm optimization (PSO) algorithm for feature selection, based on information theory. The PSO is enhanced with clustering and crowding features, which enable the algorithm to maintain a diverse set of solutions throughout the optimization process. Two objectives based on mutual information are used for selecting the optimal features, where the first aims to maximize relevance of features to the class labels, while the second to minimize the redundancy among the selected features. The proposed method is tested on four datasets, giving promising results when compared to multi-objective PSO, and multi-objective Bat algorithm.