Support Vector Machine Hyperparameter Tuning with PSO, EHO, and a Hybrid PSO–EHO Algorithm for Social Network Content Filtring

Abstract

Support Vector Machines (SVM) work especially well on high-dimensional classification tasks. However, the correct tuning of hyperparameters, particularly the kernel coefficient \( \gamma \) and regularization parameter (C), is crucial to their performance. For complicated, unbalanced, or noisy datasets, conventional tuning techniques like grid and random search frequently fail. Particle Swarm Optimization (PSO), Elephant Herding Optimization (EHO), and a hybrid PSO–EHO approach are three bio-inspired metaheuristic algorithms that are investigated in this study for SVM hyperparameter tuning. Four real-world datasets of various sizes and sampling techniques are used to assess these techniques in the context of Twitter spam detection. Twelve behavioral and content-based features are used to represent each tweet. To guarantee dependability and reproducibility, the experimental design incorporates multiple independent runs and stratified 5-fold cross-validation. PSO–EHO approach performs better, increasing F1-scores by up to 19\% when compared to untuned models, although the results demonstrate that all three algorithms significantly improve SVM classification. The strength of our approach is in fusing the diversity of solutions offered by EHO with the quick convergence of PSO. These findings demonstrate the effectiveness of hybrid bio-inspired optimization in improving traditional models, with potential applications in cybersecurity, spam detection, and social media analytics.