Learning Semantic Boundaries: An Adaptive Structural Loss for Multi-Label Contrastive Learning

Multi-Label Contrastive Learning (MLCL) seeks to pull samples with shared labels closer in an embedding space. However, existing methods primarily adjust attractive forces without explicitly shaping a geometric structure that captures complex label semantics. This work targets learning an embedding space isomorphic to the semantic label structure, a challenge complicated by ranking noise arising from dense positives and sampling distortion caused by finite queue sizes. To address these problems, we propose Hierarchical Boundary Learning (HBL), a novel structured regularization loss. HBL partitions positive samples into soft and hard subsets based on Jaccard similarity, then enforces a dual-boundary constraint: a relative boundary between soft and hard positives to mitigate ranking noise, and an absolute boundary to anchor hard positives, preventing positive sample expulsion. A reliability gating mechanism further counters sampling distortion. Experiments on diverse multi-label datasets show that MLCL methods using HBL achieve significant improvements over prior methods across multiple evaluation metrics.