Multi-label feature selection for missing labels by granular-ball based mutual information

Abstract

Multi-label feature selection serves an effective dimensionality reduction technique in the high-dimensional multi-label data. However, most feature selection methods regard the label as complete. In fact, in real-world applications, labels in a multi-label dataset may be missing due to various difficulties in collecting sufficient labels, which enables some valuable information to be overlooked and leads to an inaccurate prediction in the classification. To address these issues, a feature selection algorithm based on the granular-ball based mutual information is proposed for the multi-label data with missing labels in this paper. At first, to improve the classification ability, a label recovery model is proposed to calculate some labels, which utilizes the correlation between labels, the properties of label specific features and global common features. Secondly, to avoid computing the neighborhood radius, a granular-ball based mutual information metric for evaluating candidate features is proposed, which well fits the data distribution. Finally, the corresponding feature selection algorithm is developed for selecting a subset from the multi-label data with missing labels. Experiments on the different datasets demonstrate that compared with the state-of-the-art algorithms the proposed algorithm considerably improves the classification accuracy. The code is publicly available online at https://github.com/skylark-leo/MLMLFS.git