Feature Selection Method Based on an Improved PSO Algorithm with Multilayered Update Strategies

Feature selection (FS) is a crucial preprocessing method for improving feature set quality. Particle swarm optimization (PSO) is effective and simple to implement and has minimal parameter requirements, making it highly suitable for FS challenges. However, PSO can suffer from premature convergence and difficulty escaping local optima. To address this issue, our paper proposes a novel PSO-based feature selection technique featuring a multilevel updating strategy (MLUSPSO). This method starts with a unique classification approach in which the particle swarm is divided into three groups: elite, medium, and weak. This division is based on each particle's fitness and exploratory abilities, thereby fostering diversity in each group. Next, we define specific updating strategies for the feature subsets chosen by these particle groups. Notably, we devise a new correlation-based strategy to update weak-class particles, enhancing their exploratory potential in the feature space through the incorporation of correlation data. Comparative tests revealed that MLUSPSO outperforms other high-dimensional classification feature selection methods by delivering feature subsets with superior classification accuracy and fewer features.