Towards Noise-Robust Partial Multi-Label Learning
Zhili Qin · 2024
In many real-world scenarios, acquiring accurate label information for multi-label data is challenging, often resulting in only partial label availability. This issue defines the partial multi-label learning problem (PML), primarily complicated by significant label noise (false-positive labels). This noise introduces two novel challenges: feature-label mismatch and inaccurate label correlation. To address these issues, we introduce a new noise-robust approach for PML, termed NRPML. NRPML employs non-negative matrix factorization (NMF) to learn low-dimensional latent features and label embeddings, effectively reducing false positives and noisy features. Additionally, an orthogonal constraint on the latent label embeddings enhances local label correlation, while a linear mapping based on these embeddings facilitates final predictions. Extensive comparative experiments in both traditional and partial multi-label learning scenarios demonstrate that NRPML outperforms existing state-of-the-art algorithms.