Enhanced particle swarm optimization based on network structure for feature selection

Abstract Feature selection plays a crucial role in classification and clustering tasks, especially with the accumulation of high-dimensional data. Recently, many evolutionary algorithms have been developed to address FS problems from both single-objective or multiobjective perspectives. Among these, methods based on particle swarm optimization (PSO) have shown excellent performance due to PSO’s high efficiency and strong search capability. However, existing PSO-based methods often overlook the complex relationships among features, which contain valuable information and could enhance the search process. To address this limitation, we propose leveraging the relationship patterns hidden in the feature space to improve PSO and multiobjedctive PSO (MOPSO) for the single-objective and multiobjective FS problems. In previous work, complex feature relationships have been represented as a feature network, with important relationship patterns highlighted as structural characteristics of the feature network. Building on this, we integrate these structural characteristics into PSO and MOPSO, resulting in two network-based PSOs: NetG-PSO and NetG-MOPSO. Experimental results on 10 datasets demonstrate that the proposed NetG-PSO and NetG-MOPSO outperform baseline methods.