A Three-Stage Feature Selection Method Based on Particle Swarm Optimization with Dynamic Granularity Clustering

Multi-stage feature selection methods have attracted increasing attention due to the ability to integrate the advantages of filtering, clustering, and evolutionary algorithms. However, existing methods still face several challenges. Firstly, most feature clustering approaches require computing mutual information between features, which causes high computational cost. Secondly, fixed clustering granularity fails to adapt feature group structures based on performance feedback during the optimization process, which reduces search efficiency. To address these issues, a three-stage feature selection method based on particle swarm optimization with dynamic granularity clustering (DGCPSO-FS) is proposed. First, ReliefF, Maximal Information Coefficient, and Kullback-Leibler divergence are combined to filter out irrelevant features. Then, instead of relying on mutual information between features, relevant features are grouped based on their importance differences. Finally, a novel particle swarm optimization algorithm with dynamic granularity clustering is proposed. It removes low-quality features from poorly-performing feature groups and merges these groups with the most similar high-performing ones. The experimental results show that DGCPSO-FS could achieve better classification accuracy with a lower computational overhead compared to state-of-the-art FS algorithms on highdimensional datasets.