연구홍보포털

논문 초록 (Abstract)

Feature selection (FS) is critical in classification, aiming to identify the smallest subset of features that maximizes accuracy. Given the NP-hard nature of FS, metaheuristic algorithms (MAs) are commonly applied as effective wrapper-based FS methods. However, high-dimensional datasets with many features and limited samples pose challenges, often resulting in reduced effectiveness and increased computational costs. This study presents the Adaptive Enhanced Diversified Hiking Optimization Algorithm (AEDHOA), an improved variant of the Hiking Optimization Algorithm (HOA), crafted to address these issues efficiently. AEDHOA incorporates four key strategies: the Stratified Random Initialization Strategy (SRIS) for enhanced population diversity, the Enhanced Leader Coordination Strategy (ELCS) for multiple leader guidance to prevent premature convergence, the Adaptive Perturbation Strategy (APS) to introduce controlled randomness for escaping local optima, and the Dynamic Exploration Strategy (DES) to balance global exploration and local exploitation dynamically. AEDHOA's performance is validated using benchmark functions from CEC2017 and CEC2022, where it is compared with sets of classical, recent, and advanced algorithms to ensure comprehensive benchmarking. Additionally, AEDHOA is evaluated as a feature selection method on datasets from the UCI repository and a real-world skin cancer dataset, showcasing its capacity to overcome local minima and accelerate convergence. Comparative results reveal that AEDHOA achieves substantial improvements, with classification accuracy ranging from 0.76 to 1.00 across diverse datasets, demonstrating its robustness and effectiveness in high-dimensional FS tasks.