Contrastive active adaptive partial label learning under class distribution mismatch

Partial label learning is an important learning framework where each training sample is associated with a candidate label set and its ground-truth label is included in the candidate label set. Active partial label learning is a variation where training data consists of both labeled and unlabeled samples. However, there exists the problem of class distribution mismatch, wherein the unlabeled sample set contains many instances out of the target categories. In this paper, a contrastive active adaptive partial label learning method under class distribution mismatch which combines active partial label learning with contrastive coding is proposed. A novel active sample selection strategy is first established to use label propagation ability to measure the optimization ability of unlabeled samples to partially labeled samples. Furthermore, to solve the problem of class distribution mismatch, a joint query score based on contrastive coding is utilized to reduce the queries of unlabeled samples out of target categories. Finally, the above two indicators are combined adaptively to select the most valuable unlabeled samples in target categories for manual labeling and the selected samples will be added to the training sample set to train the new classifier. The effectiveness and efficiency of the method are evaluated by performing experiments on the datasets CIFAR10 and CIFAR100.