Confidence-Aware Relation Contrastive Learning for Partial Multi-Label Learning

Partial multi-label learning (PML) learns from over-complete candidate labels: All true labels are included in the candidate set, but irrelevant labels may also be annotated. Existing PML methods mainly reduce this ambiguity during classifier training by estimating label reliability, recovering latent label distributions, or modeling label correlations. This paper studies a complementary question: how can PML supervision be used for contrastive representation pretraining? The main difficulty is relation construction. Raw candidate label overlap is not a reliable contrastive signal, because two images can share only false candidate labels. We therefore formulate PML pretraining as a relation disambiguation problem and propose a confidence-aware relation contrastive learning framework. The method estimates candidate label confidence, uses it to construct semantic relation distributions over current and queued samples, and aligns the embedding distribution with these confidence-weighted relations. The same candidate label confidence semantics are used during downstream fine-tuning, keeping non-candidate labels as reliable negatives while softly disambiguating the candidate set. Across VOC2007, NUS-WIDE, and COCO2014, the method improves subset accuracy and F1-oriented measures under moderate and severe partial-label noise. For example, on VOC2007 at p=0.8, it improves subset accuracy from 0.3322 to 0.5269 and micro-F1 from 0.6710 to 0.7617 over PML-CD. However, the CUB200 results further identify a limitation: fine-grained attribute prediction is less consistently served by the present global representation pretraining design.