An Improved Multi-objective Particle Swarm Optimization with Adaptive Penalty Value for Feature Selection

Abstract

Feature selection is an important data-preprocessing technique to eliminate the features with low contributions in classification. Currently, many researches focus their interests on the combination of feature selection and multi-objective particle swarm optimization (PSO). However, these methods exist the problems of large search space and the loss of global search. This paper proposes a multi-objective particle swarm optimization with the method called APPSOFS that the leader archive is updated by an adaptive penalty value mechanism based on PBI parameter. Meanwhile, the random generalized opposition-based learning point (GOBL-R) point is adopted to help jump out of local optima. The proposed method is compared with three multi-objective PSO and MOEA/D on six benchmark datasets. The results have demonstrated that the proposed method has better performance on feature selection.