Can Label-Specific Features Help Partial-Label Learning?

Partial label learning (PLL) aims to learn from inexact data annotations where each training example is associated with a coarse candidate set. Due its practicability, many PLL algorithms have been proposed in recent literature. Most prior works attempt identify the ground-truth labels sets and classifier trained afterward by fitting features of examples their exact labels. From different perspective, we propose enrich feature space raise question ``Can label-specific help PLL?'' rather than identical for all classes. Despite benefits, previous approaches rely on split positive negative class then conduct clustering analysis, which not directly applicable PLL. To remedy this problem, an uncertainty-aware confidence region accommodate false We first employ graph-based enhancement yield smooth pseudo-labels facilitate split. After acquiring features, family binary classifiers induced. Extensive experiments both synthesized real-world datasets are conducted results show that our method consistently outperforms eight baselines. Our code released at https://github.com/meteoseeker/UCL