Reduction from Complementary-Label Learning to Probability Estimates

Complementary-Label Learning (CLL) is a weakly-supervised learning problem that aims to learn multi-class classifier from only complementary labels, which indicate class an instance does not belong. Existing approaches mainly adopt the paradigm of reduction ordinary classification, applies specific transformations and surrogate losses connect CLL back classification. Those approaches, however, face several limitations, such as tendency overfit. In this paper, we sidestep those limitations with novel perspective–reduction probability estimates classes. We prove accurate labels lead good classifiers through simple decoding step. The proof establishes framework estimates. offers explanations key its special cases allows us design improved algorithm more robust in noisy environments. also suggests validation procedure based on quality estimates, offering way validate models CLs. flexible opens wide range unexplored opportunities using deep non-deep for solve CLL. Empirical experiments further verified framework’s efficacy robustness various settings. full paper can be accessed at https://arxiv.org/abs/2209.09500 .