<p>Feature selection aims to identify the most relevant and non-redundant features from high-dimensional data, which is challenging due to the large number of possible feature combinations. This has motivated the use of metaheuristic optimization techniques. Among these, Binary Particle Swarm Optimization (BPSO) is frequently used for its simplicity and computational efficiency. However, BPSO often suffers from premature convergence and stagnation in local optima. These issues arise from several factors, including uninformed random initialization, fixed inertia weight, suboptimal velocity-to-binary mapping, and reliance on a single global best solution. To address these challenges, we propose an unsupervised hybrid method that integrates the Scalable Laplacian score with an enhanced BPSO framework. The main contributions of this work are as follows: (i) a hybrid initialization strategy guided by Laplacian scores to improve initial population quality and reduce excessive randomness; (ii) a scalable computation of Laplacian scores for efficient processing of large-scale high-dimensional datasets; (iii) a sub-population division mechanism with iteration-wise random assignment of inertia weight schemes to improve the trade-off between exploration and exploitation; (iv) a V-shaped transfer function for effective velocity-to-binary mapping in feature subset selection; (v) a fitness evaluation strategy using the Silhouette Index through K-means clustering, coupled with a fitness caching mechanism to reduce redundant computations and improve computational efficiency; and (vi) the proposed algorithm simultaneously maintains species-level best solutions alongside the global best solution, improving robustness and mitigating stagnation. Experimental results on 15 benchmark and real-life plant genomics datasets demonstrate that the proposed method achieves superior clustering quality and computational efficiency compared with state-of-the-art feature selection techniques.</p>