Partial Multi-label Learning with Decorrelated Label Encoding and Class Prototype Decoding

In partial multi-label learning (PML), each instance corresponds to candidate labels containing both ground-truth and noisy labels. Its core task is to achieve accurate label denoising and classifier learning under inaccurate supervision. However, existing methods suffer from: 1) ignoring noise-caused label association distortion, where directly learning correlations from labels with noise leads to unreliable guidance information; 2) the inability to directly adapt class prototype learning to PML tasks, as existing prototype methods cannot handle the discrete multi-label nature conflicting with probabilistic modeling requirements. To address these issues, we propose a novel partial multi-label learning method with decorrelated label encoding and class prototype decoding (PML-LED). First, we propose a label-based encoding-decoding framework that encodes candidate labels into orthogonal latent space to eliminate label associations, then decodes them into purified labels and uses ℓ₁ norm for capturing sparse differences between purified and original labels. Second, to overcome the problem that class prototype learning cannot be directly applied, we normalize label values into probability distributions compatible with fuzzy membership semantics via softmax function, enabling effective feature guidance for the above process. Meanwhile, a class prototype decoding mechanism is proposed to provide consistent guidance for the encoding and decoding processes, achieving correct label association recovery. Finally, an ℓ_2,1 norm regularized classifier achieves feature selection. Extensive experiments demonstrate that PML-LED outperforms state-of-the-art methods.