Partial Multi-label Learning via Label Anchor Graph

Partial Multi-Label Learning (PML) addresses scenarios where each instance is associated with a candidate label set containing both relevant and irrelevant labels. The coexistence of noisy and missing labels poses significant challenges for reliable classifier learning and label disambiguation. Existing approaches often neglect the exploitation of co-occurring label structures and suffer from inaccurate feature-to-label mappings caused by linear projection assumptions. This paper presents Partial Multi-Label Learning via Label Anchor Graph (PML-LAG), a dynamic anchor-based framework that jointly refines semantic prototypes and classifier mappings under dual manifold constraints. Dynamic label anchors construct a semantic prototype graph capturing frequent label co-occurrences, while an anchor Laplacian preserves semantic smoothness among prototypes. In parallel, a neighborhood-preserving embedding enforces local feature consistency, and a collaborative optimization scheme aligns the feature and prototype spaces for robust label reconstruction. Extensive experiments on multiple real-world and synthetic datasets demonstrate that PML-LAG achieves superior noise robustness and label disambiguation performance compared with state-of-the-art partial multi-label learning methods.