RaCo: Reliable multi-augmentation complementary label learning

Complementary Label Learning (CLL) represents a fundamental weakly-supervised learning paradigm where training instances are annotated with complementary labels - each indicating a class to which the instance does not belong. While existing approaches have attempted to incorporate advanced augmentation alignment techniques to overcome the challenges of this extremely weak supervision, they fail to address a critical aspect: How to effectively leverage information from multiple augmented views to guide model learning. In this paper, we present the first integration of Evidential Deep Learning (EDL) into CLL, introducing a principled framework to quantify and utilize view-specific uncertainty in CLL. Specifically, we propose a novel framework termed RaCo, i.e., Reliable Multi-augmentation Complementary Label Learning, which leverages EDL to estimate second-order uncertainty for each augmented view and then calculate a stable metric, named augmented view certainty (AVC), to assess the quality of each view. Using the proposed AVC, RaCo formulates a reliable target distribution by combining the predictions of multiple augmented views and further employs Kullback-Leibler (KL) divergence to align the prediction probability of each augmented view with this reliable distribution. Extensive experiments on the benchmark datasets under various CLL settings validate the effectiveness and superiority of the proposed RaCo. Our code will be released upon acceptance.