Multi-label feature selection based on rough granular-ball and label distribution

The explosive growth of datasets is always accompanied by dimension disasters, which have become more common in multi-label data. Various feature selection techniques are devised to obtain a compact and discriminative feature subset. As an efficient and robust model of granular computing, granular-ball computing delivers a novel technology for data granulation, imitating the characteristic of large-scale priority of human cognition. In this paper, an effective model of granular-ball computing called rough granular-ball computing is proposed, which clusters multi-label data into multiple granules that reflect the local information of instances, and a label enhancement approach is proposed to convert logical labels into label distribution by exploring the similarity between the instance and the rough granular-ball. Moreover, a rough granular-ball-based feature selection method is proposed for enhanced label distribution data, which measures the feature significance through labels' consistency of samples within the same information granularity. Finally, the proposed method is compared with seven state-of-the-art multi-label feature selection algorithms on twenty two public benchmark datasets in terms of six evaluation metrics. The experimental results show that our proposed methods obtain a superior classification performance.