Dual Memory Matrix Gannet Optimization with Dynamic Weight Adjustment for Robust Feature Selection

Feature selection tasks are commonly handled by swarm intelligence algorithms, but they often struggle to maintain robustness across diverse datasets. The Gannet Optimization Algorithm (GOA) provides quick solutions for complex problems and performs well on constrained engineering design problems. In spite of this, it faces challenges in exploration due to its U-shape and V-shape dive patterns. We propose the Improved Gannet Optimization Algorithm (IGOA), which enhances GOA's performance through the introduction of two population matrices, thus striking a better balance between exploration and exploitation. To prevent the algorithm from getting stuck in local optima, two memory matrices are maintained in IGOA. These matrices increase the strength of high-quality solutions and allow exploration of new areas of the search space. Based on performance metrics, a fuzzy-driven fitness function adjusts weights assigned to classification accuracy, feature ratio, Cohen's kappa, and Matthews correlation coefficient dynamically. Fuzzy systems optimize fitness functions using 135 rules, adapting weights for a more nuanced evaluation of solutions. Several state-of-the-art algorithms are compared with IGOA on 17 datasets ranging from 4 to 1203 features, including the Genetic Algorithm (GA), the Whale Optimization Algorithm (WOA), and the Elephant Herding Optimization (EHO). Based on the results, IGOA significantly outperforms GOA, achieving increases in precision and recall metrics and a reduction in mean fitness value by 71%, especially in high-dimensional scenarios. In addition, IGOA selects fewer features while maintaining classification accuracy. Support Vector Machines (SVMs) are used to validate the method’s effectiveness across datasets and feature selection challenges. Additionally, IGOA is compared to both optimization and machine learning methods for accuracy, feature reduction, and time complexity, making it a compelling solution for real-world feature selection.