Abstract
Abstract
Enhancing the classification performance is more challenging than reducing the number of features in multiobjective feature selection. Such imbalance might lead the population to converge to a region on the Pareto front with a small number of features. Meanwhile, increasing the classification performance is time consuming, especially when the training data contain a large number of instances and features. Surrogate models can help reduce the optimization time. However, they need to be well-designed to accurately approximate such complex search space. Over-reliance on surrogate models, especially those with large approximation errors, can mislead the population toward local optima regions, or even impede population convergence. To address the above issues, we propose a preference-driven surrogate-assisted evolutionary multiobjective feature selection algorithm, where three surrogate-based strategies are developed by introducing objective preferences. A surrogate-assisted initialization strategy is employed to initialize a set of high-quality feature subsets that encompass important features. Two surrogate-assisted evolutionary search strategies are tailored to effectively identify feature subsets with a small number of features and feature subsets with high classification performance with the assistance of single and multiple surrogate models, respectively. Moreover, the two surrogate-assisted evolutionary search strategies are combined with the full-evaluation-based search to further increase the classification performance. The proposed method has been tested on 23 diverse datasets. The experimental results demonstrate that the performance of the proposed algorithm outperforms that of six other recently developed methods under the same time budget.