Selective Disambiguation and Ranking-Aware Distillation for Partial Multi-Label Learning

Partial multi-label learning (PML) aims to train a predictive model from data where each instance is associated with a set of candidate labels, containing both relevant ground-truth labels and irrelevant noisy ones. Existing PML methods seldom account for the heterogeneous difficulty of disambiguation among samples. Their uniform treatment of both easy and highly ambiguous samples frequently limits model effectiveness. To address this limitation, we propose Selective Disambiguation and Ranking-Aware Distillation for Partial Multi-Label Learning (SR-PML), a novel two-stage framework that enhances PML by leveraging sample selection. In the first stage, we develop an enhanced Expectation-Maximization (EM) algorithm that refines label confidences via a distance-weighted local consensus mechanism, adaptively isolating a highly reliable confident subset of distinct samples using a normalized entropy criterion. To safely transfer this purified knowledge to the remaining ambiguous samples without introducing early-stage confirmation bias, the second stage introduces a structurally-aware teacher-student distillation mechanism. The teacher model, securely trained exclusively on the anchor set, generates high-fidelity pseudo-labels governed by a novel Confidence-Rank Alignment (CRA) and variance regularization loss. This synergy explicitly preserves fragile label ranking structures and prevents noise accumulation, enabling the student model to robustly generalize across the entire dataset. Extensive experiments on 36 synthetic and real-world PML datasets show that SR-PML substantially outperforms state-of-the-art methods.