Enhancing Partial Label Learning via Multi-Task Contrastive Learning and Prototype-Guided Disambiguation

Partial Label Learning (PLL) is an important weakly supervised learning framework, with its core being that each training sample is associated with a candidate label set consisting of true labels and noisy labels. Although PLL is highly suitable for real-world data annotation scenarios with label ambiguity, it faces two major challenges: representation learning and label disambiguation. Existing work addresses representation learning or label disambiguation separately, but they are limited by the constrained capacity of representation learning and cannot effectively align with the task of label disambiguation. To overcome this difficulty, this paper proposes a novel multi-task framework that employs a contrastive learning module for multi-task collaboration, enabling the model to learn good representations and incorporating prototypes into the classification loss to assist in label disambiguation. Extensive experiments demonstrate that the performance of our proposed method surpasses that of the current most advanced methods.