Abstract: 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 ...