Multi-Label Feature Selection with Granular-Ball Multi-Scale Modeling and Label Distribution Learning

Traditional single-label learning assumes each instance belongs to only one category, which limits its ability to describe real-world objects with multiple semantics. Although multi-label learning enables multiple label assignments, its binary label representation cannot capture label significance differences. Label distribution learning addresses this by assigning description degrees to form label probability distributions. To handle high-dimensional multi-label data, this paper proposes a granular-ball based multi-scale feature selection method integrating the principle of large-scale first in granular-ball computing with label distribution learning (GBMSLD). This approach constructs multi-scale information granules under the label distribution framework and integrates cross-scale uncertainty measurements for feature screening. Experiments on multiple datasets show that the performance of the proposed method performs well in four evaluation metrics.