Dual feature-driven approach for partial multi-label learning

Partial multi-label learning (PML) addresses the challenge of instances annotated with multiple candidate labels, of which only a subset are valid. Existing PML methods primarily focus on label disambiguation through label relevance or label propagation techniques, yet overlook the critical impact of ambiguous features that may misguide model training. This paper first establishes a generalization error bound via Rademacher complexity, theoretically proving that ambiguous features linearly increase model complexity. Motivated by this, we propose a Dual Feature-driven PML method (DF-PML), which introduces two key innovations: (1) a noise feature identification mechanism based on low-rank and sparse assumptions combined with matrix decomposition techniques. It aims to eliminate the interference of noisy features, which is crucial since the presence of noisy features can obscure the identification of ambiguous features; and (2) an ambiguous feature recognition mechanism. It employs a shallow network to estimate the ambiguity probability of each feature and incorporates label correlations and neighborhood granularity to compute label confidence. Consequently, the impact of ambiguous features is mitigated, leading to improved model performance. Experimental results show that the proposed DF-PML method outperforms existing methods regarding overall performance.