An improved feature selection method based on angle-guided multi-objective PSO and feature-label mutual information

Multi-objective particle swarm optimization (MOPSO) has been widely applied to feature selection. Although these MOPSO-based feature selection methods have achieved good performance, they still need improvement in obtaining high-quality feature subsets. Most feature selection methods mainly consider improving the classification accuracy and reducing the number of selected features while ignoring some prior information in the feature data, resulting in weak interpretability of the finally selected features. Moreover, the MOPSO algorithm itself still has some defects. With the increase in the feature dimension, PSO easily stagnates in local minima due to the lack of sufficient selection pressure. In this paper, an improved feature selection method based on angle-guided MOPSO and feature-label mutual information is proposed to select high-quality feature subsets. On the one hand, to select the features that are more related to the class labels, we propose an adaptive threshold setting strategy based on feature-label mutual information. This strategy extracts useful prior information from the original data and encodes it throughout the entire feature selection process. The introduction of mutual information helps to enhance the interpretability of the selected features. On the other hand, to improve the performance of MOPSO, a global leader selection strategy based on the minimum angular distance information is proposed to guide the swarm to converge to Pareto front. The proposed method is compared with six multi-objective feature selection methods on six UCI benchmark datasets. Experimental results verify that the proposed algorithm could achieve satisfactory results in terms of both improving classification accuracy and the interpretability as well as reducing the number of selected features.