dual-phase local search embedded multi-verse optimizer for optimal feature subset selection

Feature selection plays a pivotal role in enhancing the performance of machine learning models by reducing dimensionality, improving interpretability, and minimizing computational overhead. this study presents the improved multi-verse optimizer (imvo), a new feature selection algorithm that merges the global exploration ability of the standard multi-verse optimizer (mvo) with a dual-phase mutation-based local search mechanism. unlike previous mvo-based or hybrid metaheuristic approaches, imvo simultaneously strengthens exploitation through targeted refinement of the best solutions and preserves population diversity via periodic random-solution mutations. this strategic combination mitigates premature convergence, accelerates convergence speed, and improves robustness across diverse high-dimensional datasets. comprehensive experiments on 14 widely used datasets obtained from the uci repository show that imvo consistently achieves higher classification accuracy, fewer selected features, and lower fitness values than five state-of-the-art algorithms (mvo, ga, pso, ssa, hho). quantitative analysis using the wilcoxon signed-rank test certifies the significance of these enhancements, underscoring the algorithm’s reliability. while the inclusion of local search increases computational cost, the demonstrated gains in accuracy, stability, and feature reduction affirm this cost-benefit relationship, positioning imvo as a competitive and versatile tool for feature selection and related optimization problems.