A Particle Swarm Optimizer with Biased Exploration and Exploitation for High-Dimensional Optimization and Feature Selection

Autonomous intelligent systems have been widely implemented in a broad range of applications. These applications often involve data-dense tasks where an efficient feature selection process is required to eliminate redundant features and improve model performance. However, the feature selection tasks with high dimensionality still remain challenging to deal with. In order to address high-dimensional feature selection problems with greater effectiveness and efficiency, this paper proposes a particle swarm optimizer variant named PSO-BEE that allows the swarm to take full advantage of exploration and exploitation at due evolutionary stages. At the early stage, a large size of candidate exemplar group is formed for diversity enhancement, attempting to find as many new combinations of features as possible. At the later stage, a tiny size of candidate exemplar group is constructed, allowing updated particles to learn from the few best elite exemplars to continuously refine feature selection solutions with a lower classification error rate. Three PSO-BEE variants with distinct preferences for exploration and exploitation are proposed based on different parameter adjustment strategies. Experiments are executed on 500, 1000, and 2000-dimensional benchmark suites presented by CEC and real-world feature selection datasets from the Machine Learning Repository. Experimental results demonstrate the competitive performance of PSO-BEE in high-dimensional global optimization and feature selection when compared with several state-of-the-art approaches.