Multi-objective feature selection using PSO with rough hypercuboid approach

Feature selection aims to select a minimal subset that maintains or improves classification accuracy. Incorporating rough set theory with the hypercuboid equivalence partition matrix, the rough hypercuboid (RH) approach serves as a criterion for selecting significant features from high-dimensional datasets. However, when integrating RH with meta-heuristic optimization algorithms, the focus remains on maximizing classification performance, lacking a single-objective mechanism to simultaneously minimize subset size. To overcome this, we propose a novel criterion using Particle Swarm Optimization (PSO) with an efficient RH-based fitness function that simultaneously considers maximizing classification performance and minimizing subset size, namely RHPSON. We employ RH relevance criteria for particle initialization and significance measure to reduce class imbalance. Results indicate that RHPSON outperforms representative algorithms in classification accuracy, subset compactness, and efficiency. Addressing single-objective limitations, this paper introduces a multi-objective algorithm that generates a Pareto front of nondominated solutions in classification accuracy and subset size, called the nondominated sorting based feature selection using PSO with an efficient RH fitness function (NSRHPSO). It improves gbest selection by integrating thermodynamic probability to balance convergence to Pareto front and population diversity. Experimental results demonstrate that NSRHPSO achieves fewer features with competitive accuracy than representative multi-objective algorithms.