Multi-Label Feature Selection by Maximum Affinity-Relevance, Maximum Affinity-Interaction and Minimum Affinity-Redundancy

Multi-label feature selection solves the high-dimensional challenge problem in multi-label learning, and is widely used in pattern recognition, machine learning, and other related fields. However, most existing multi-label learning methods neglect the affinity between the features and the labels, which means the selected features may only provide categorization information for some of the labels rather than for all of them, resulting in one portion of labels obtaining sufficient categorization information, while another portion of labels obtains insufficient categorization information. In order to address this issue, we propose a new method for multi-label feature selection. Firstly, we propose the notion of affinity, which states that features are biased to provide categorical information for labels. Secondly, we propose three new feature evaluation metrics, affinity-relevance, affinity-redundancy, and affinity-interaction, which evaluate features in three dimensions from an affinity perspective. Thirdly, a new feature evaluation function called feature affinity significance, as well as a feature selection algorithm based on MARAImAR are proposed. Finally, experimental results on 15 multi-label benchmark datasets demonstrate that the proposed method outperforms the other nine representative multi-label feature selection methods.