Hyperparameter Tuning of SVM Using Metaheuristics: An Empirical Analysis
Sunday Oladayo Oladejo, Lateef Adesola Akinyemi, Stephen Obono Ekwe, Taiwo Gabriel Omomule, Oluwaseyi Paul Babalola · 2025
The Support Vector Machine (SVM) performance is highly dependent on the selection of optimal hyperparameterse Conventional optimisation methods, including grid search, gradient descent, randomized search, and various experimental approaches, have been extensively utilized in the literature to identify optimal hyperparameters. This paper provides an empirical analysis of SVM's hyperparameter tuning by use of metaheuristics. We compare 11 widely known metaheuristics, such as the Genetic Algorithm (GA), Ant Colony Optimisation (ACO), Particle Swarm Optimisation (PSO), and Differential Evolution (DE), and employ three common datasets, namely Iris, Breast Cancer Wisconsin, and MNIST, due to their unique characteristics. Empirical analysis takes into account factors including cross-validation, number of agents, noise, and mis-labelled data insertion, with emphasis on performance metrics such as accuracy, number of support vectors, and computational time. Our findings reveal that the PSO outperforms other metaheuristics, followed by the DE and ACO. In addition, we conclude that the ACO has the least computational time.