Abstract

Feature selection (FS) is crucial in supervised learning, mainly when dealing with high-dimensional datasets, since the models’ efficiency decreases due to the curse of dimensionality. Reducing the number of features enhances computational efficiency and improves models’ interpretability and generalization. Support vector machine (SVM) has been widely used with FS due to its ability to assign importance to individual features through feature-specific regressors, which can be eliminated when deemed irrelevant. This paper proposes a multi-objective metaheuristic approach based on non-dominated sorting genetic algorithm II, integrating FS into the soft-margin SVM model to optimize both predictive performance and computational efficiency. Unlike prior methods with static FS, our approach dynamically selects features to approximate the Pareto-optimal frontier, balancing structural and empirical risk. The proposed algorithm incorporates a novel solution representation, specialized crossover and mutation operators, and a weighted optimization strategy to effectively handle dataset imbalances. Additionally, we apply effective parameter tuning based on three considered performance metrics, resulting in three distinct versions of our approach, each exhibiting different search behaviors. Extensive experiments on well-known binary classification datasets demonstrate that all three versions outperform the state-of-the-art algorithm in both predictive performance and computational efficiency within the given time limits. Among them, the version that employs a conservative FS strategy, maintaining larger feature subsets while applying high mutation rates, achieved the best overall results. In addition, our approach also exhibits competitive performance on real-world large-scale datasets. © 2025 Elsevier B.V.