A novel multi-swarm particle swarm optimization for feature selection

A novel feature selection method based on a multi-swarm particle swarm optimization (MSPSO) is proposed in this paper. The canonical particle swarm optimization (PSO) has been widely used for feature selection problems. However, PSO suffers from stagnation in local optimal solutions and premature convergence in complex feature selection problems. This paper employs the multi-swarm topology in which the population is split into several small-sized sub-swarms. Particles in each sub-swarm update their positions with the guidance of the local best particle in its own sub-swarm. In order to promote information exchange among the sub-swarms, an elite learning strategy is introduced in which the elite particles in each sub-swarm learn from the useful information found by other sub-swarms. Moreover, a local search operator is proposed to improve the exploitation ability of each sub-swarm. MSPSO is able to improve the population diversity and better explore the entire feature space. The performance of the proposed method is compared with six PSO based wrappers, three traditional wrappers, and three popular filters on eleven datasets. Experimental results verify that MSPSO can find feature subsets with high classification accuracies and smaller numbers of features. The analysis of the search behavior of MSPSO demonstrates its effectiveness on maintaining population diversity and finding better feature subsets. The statistical test demonstrates that the superiority of MSPSO over other methods is significant.