Feature Selection Using PSO: A Multi Objective Approach

Abstract

Feature selection is a pre-processing technique in which a subset or a small number of features, which are relevant and non-redundant, are selected for better classification performance. Multi-objective optimization is applied in the fields where finest decisions need to be taken in presence of trade-offs between two or more differing objectives. Therefore, feature selection is considered as a multi-objective problem with conflicting measures like classification error rate and feature reduction rate. The existing algorithms, Non-dominated Sorting based particle swarm optimization for Feature Selection (NSPSOFS) and Crowding Mutation Dominance based particle swarm optimization for Feature Selection (CMDPSOFS) are the two multi-objective PSO algorithms for feature selection. This work presents the enhanced form of NSPSOFS and CMDPSOFS. A novel selection mechanism for gbest is incorporated and hybrid mutation is also added to the algorithms in order to generate a better pareto optimal front of non-dominated solutions. The experimental results show that the proposed algorithm generates non-dominated solutions and produce better result than existing algorithms.