Enhanced Multi-Instance Partial Label Learning via Average Gradient Outer Product

Multi-instance partial-label learning (MIPL) is a recently proposed learning paradigm to address tasks that multi-instance bags are associated with a candidate label set comprising one ground-truth label and several false positive labels. Existing MIPL methods rely on simple instance level information, and can hardly find the key instances under noisy labels. In this paper, we propose a novel algorithm termed AGOPMIPL, i.e., Average Gradient Outer Product based Multi-Instance Partial-Label Learning to address the problem. AGOP derives a data-dependent metric in the embedding space by computing the outer product of classifier gradients, which stretches discriminative feature dimensions and facilitates more accurate key instance identification. Moreover, AGOP aggregates gradient information across all training samples, providing inherent robustness to label noise. Additionally, we introduce a progressive label disambiguation strategy that gradually refine the learning targets. Experimental studies on benchmark and real-world datasets demonstrate the superiority of AGOPMIPL over existing MIPL methods.