Learning with Label Proportions Based on Adaptive Confidence and Instance Similarity

Learning with Label Proportions (LLP) copes with the weakly supervised problem that training instances are grouped as bags, among which only the proportions of different classes in each bag are provided, leaving the instance-level class labels inaccessible. Most LLP research focuses on finding instance-level labels, ignoring disambiguation difficulty and relationships between instances. This paper introduces LLP-ACIS, a model that tackles these issues. First, random augmentations are adopted to enlarge the diversity of original datasets to improve the generalization ability. Next, we select high-confident instances by imposing adaptive thresholds on the predicted outputs of weakly augmented instances. Finally, by introducing the similarity loss, we pull closer the high-confident weakly augmented instances with their similar strongly augmented counterparts in the output space. Empirical experiments conducted on four benchmark image classification datasets justify the superiority of LLP-ACIS in prediction accuracies.