Partial-Label Regression

Partial-label learning is a popular weakly supervised setting that allows each training example to be annotated with set of candidate labels. Previous studies on partial-label only focused the classification where labels are all discrete, which cannot handle continuous real values. In this paper, we provide first attempt investigate regression, real-valued To solve problem, propose simple baseline method takes average loss incurred by as predictive loss. The drawback lies in true label may overwhelmed other false overcome drawback, an identification least We further improve it proposing progressive differentiate using progressively updated weights for losses. prove latter two methods model-consistent and convergence analysis showing optimal parametric rate. Our proposed theoretically grounded can compatible any models, optimizers, Experiments validate effectiveness our methods.