Structured insight: an innovative disambiguation paradigm for semi-supervised partial label learning

Abstract

Abstract Semi-Supervised Partial Label Learning (SPLL) aims to learn from both partial label data where each instance is associated with a candidate label set and unlabeled data. Most SPLL methods work by generating pseudo-candidate labels for unsupervised data. Since the pseudo candidate labels are assigned manually to unlabeled data, sometimes they are not as accurate as the candidate labels of partial label data. However, many SPLL methods adopted unique disambiguation strategy to tackle both partial label data and unlabeled data, which often fail to account for the intrinsic structural differences between partial label data and unlabeled data, resulting in suboptimal disambiguation and degraded performance. To tackle this problem, we propose a Structured Insight-driven Disambiguation Paradigm (SIDP). SIDP disambiguates partial label data via feature-label consistency, whereas employs label expansion to generate pseudo candidate labels for unlabeled data, followed by a refinement through class perception. At last, SIDP generates the final confidence matrix by integrating the above processes via label propagation. Extensive experiments on real-world datasets demonstrate that SIDP significantly outperforms several state-of-the-art competing methods in terms of classification accuracy.