Multi-Label Feature Selection Based on Multi-Scale Fuzzy–Rough Sets

In recent years, multi-label learning has been widely applied across various domains. As a crucial preprocessing technique, multi-label feature selection significantly enhances classification performance by identifying informative features. However, real-world multi-label data often exhibit both ambiguity and uncertainty in labels, while natural phenomena tend to display complex patterns and relationships across multiple scales. Despite such complexity, existing multi-label feature selection algorithms still fall short in effectively leveraging multi-scale information, posing considerable challenges.To address these issues, this paper extends multi-scale learning to the multi-label setting and proposes a novel Multi-label Feature Selection approach based on Multi-Scale fused fuzzy uncertainty measures, termed (MSLRFS). Specifically, multi-label multi-scale fuzzy entropy is defined to integrate granular information across different scales, which is then effectively incorporated into the feature selection process. First, multi-label multi-scale fuzzy mutual information is employed to capture the maximum relevance between features and labels. Second, a stepwise feature selection strategy is proposed, which comprehensively considers feature redundancy, the relevance between features and labels under label redundancy, and their complementary nature. Furthermore, a feature selection algorithm based on a multi-label multi-scale fuzzy entropy evaluation metric is designed.In the experimental section, the proposed MSLRFS method is compared with several state-of-the-art approaches. Results demonstrate the effectiveness of MSLRFS across nine multi-label datasets, showing significant improvement in the performance of multi-label classification tasks.