Hybrid Butterfly-Grey Wolf Optimization (HB-GWO): A Novel Metaheuristic Approach for Feature Selection in High-Dimensional Data

Abstract

Feature selection plays a pivotal role in high-dimensional data analysis by reducing model complexity, improving generalization, and enhancing interpretability. This paper introduces Hybrid Butterfly-Grey Wolf Optimization (HB-GWO), a novel metaheuristic that fuses the global exploration capacity of the Butterfly Optimization Algorithm (BOA) with the local exploitation strength of the Grey Wolf Optimizer (GWO) through an adaptive exponential switching mechanism. The algorithm is designed to dynamically adjust exploration and exploitation phases over time, driven by a theoretically justified decay function. Extensive experiments were conducted on both benchmark datasets (e.g., Madelon, Colon Cancer, Arrhythmia) and a real-world high-dimensional RNA-seq dataset containing over 120,000 features, using multiple classifiers including Random Forest, SVM, XGBoost, and MLP. Results demonstrate that HB-GWO consistently outperforms classical (GA, PSO) and recent hybrid methods (Spider Wasp Optimization, Puma Optimizer), achieving superior performance in classification accuracy, AUC, F1 score, and feature reduction. Statistical tests, feature stability analysis (Jaccard index), and multi-objective extensions (Pareto front analysis) further validate its robustness. The full implementation, parameter settings, and reproducibility toolkit are released as open source.