A New Feature Selection Algorithm Based on Adversarial Learning for Solving Classification Problems

As a data preprocessing technique, the objective of feature selection (FS) is to eliminate irrelevant and redundant features. In recent years, evolutionary computation (EC) has shown great potential in solving FS problems. However, most of the existing EC-based FS approaches still face some shortcomings. To address these issues, a novel FS algorithm based on adversarial learning is proposed in this paper. In this work, an inter-population adversarial learning strategy is proposed. It enables two subswarms to compete and learn from each other in the evolutionary process. In addition, an interval-based flipping strategy is proposed. In this strategy, some features of a certain interval are flipped. A novel surrogate model is employed to pre-evaluate the candidate solutions, aiming to reduce the computational cost. Experiments on 10 UCI datasets indicate that the proposed SSAPSO is able to select smaller subsets of features in most cases with high classification accuracy.