Reflective Learning-Based Particle Swarm Optimization for Feature Selection

Swarm intelligence (SI) algorithms are widely used in feature selection (FS) problems. Yet, they often get trapped in local optima and have limited search capability. To address these issues, this paper proposes a reflective learning-based particle swarm optimization (PSO) algorithm named RPSO for FS problems. The main goal is to develop an efficient FS method by using domain knowledge and historical update information. Initially, a novel hybrid initialization is designed, incorporating the maximal information coefficient (MIC), ReliefF, and random initialization to generate a high-quality initial population. An exploration and exploitation switch strategy accelerates convergence and prevents stagnation. An adaptive movement strategy improves particle updates by considering feature importance, thereby improving adaptability. Finally, a reflective learning strategy utilizes the historical update information of particles to guide particle movement, thereby enhancing algorithmic reliability. Experimental results on 20 benchmark datasets demonstrate that RPSO outperforms the original PSO and other comparison algorithms across key performance metrics, validating the effectiveness of the four designed strategies. Thus, RPSO is a promising SI algorithm for FS problems.