An Improved Particle Swarm Optimization via Velocity-Based Reinitialization for Feature Selection

Abstract

The performance of feature selection method is typically measured based on the accuracy and the number of selected features. The use of particle swarm optimization (PSO) as the feature selection method was found to be competitive than its optimization counterpart. However, the standard PSO algorithm suffers from premature convergence, a condition whereby PSO tends to get trapped in a local optimum that prevents it from being converged to a better position. This paper attempts to improve the velocity-based initialization (VBR) method on the feature selection problem using support vector machine classifier following the wrapper method strategy. Five benchmark datasets were used to implement the method. The results were analyzed based on classifier performance and the selected number of features. It was found that on average, the accuracy of the particle swarm optimization with an improved velocity-based initialization method is higher than the existing VBR method and generally generates a lesser number of features.