Dual-Strategy Optimization of SVM Using Improved Whale Optimization Algorithm for Multi-Class Data Mining

Abstract

To address the challenges of high computational complexity and limited generalization ability in traditional support vector machines (SVM) for large-scale and multiclass datasets, this study proposes an SVM optimization model integrating an Improved Whale Optimization Algorithm (IWOA) and dual strategies. Specifically, a nonlinearly decreasing convergence factor and adaptive inertia weight are introduced to enhance the global and local search capabilities of IWOA. A redundant sample removal strategy based on K-means clustering and Fisher projection is designed to filter low-value training data. Furthermore, a distributed parallel optimization strategy with time feedback is introduced to balance node load and improve optimization efficiency. The experimental results on several public datasets (Iris, Wine, CIFAR-10, and Fashion-MNIST) demonstrated that the proposed model outperformed other benchmark algorithms, achieving the highest classification accuracy of 97.8%. In addition, the model achieved a minimum classification error of only 0.15 on the test set, significantly lower than the other three comparison models. Therefore, by incorporating the IWOA and dual strategies, the proposed model effectively enhances the classification accuracy and computational efficiency of SVM in large-scale multiclass data mining tasks.