Augmenting Graph-Based Partial Label Learning with Predictive Representations

Partial label learning (PLL) addresses the problem of learning from training examples with candidate label sets, where only one label is ground-truth. The key challenge lies in disambiguating the candidate labels while learning an accurate classifier. Graph-based PLL methods have emerged as a mainstream paradigm, which leverage the manifold structure of the feature space to propagate disambiguation information across neighboring instances. However, these methods construct the disambiguation graph exclusively from the original feature space, which may be corrupted by noise, outliers, and class-irrelevant variations. In this paper, we propose a novel framework termed Predictive Representation Augmentation (PRA) to address this limitation. The key idea is to first train a base graph-based PLL method on the original data, then use the predictive representations produced by the trained model as enhanced label-space information to guide graph construction, and augment the prediction model input with the concatenation of original features and predictive representations for retraining. In this way, the disambiguation graph benefits from structural information derived from both the feature space and the enhanced label space. We instantiate our framework on three representative graph-based PLL methods, namely PL-LEAF, PL-AGGD, and PL-CL. The experimental results on six real-world datasets demonstrate that PRA consistently improves the classification performance of all three base methods and, after Holm–Bonferroni correction, achieves statistically significant improvements in 17 out of 18 pairwise comparisons.