Imbalance Label Enhancement with Adaptive Neighbors

Label distribution learning (LDL) is gaining popularity due to its more generalized capacity to deal with label ambiguity. The cost of collecting label distribution information for data, as opposed to logical labels, is significantly higher. Thus, it is suggested to apply label enhancement to extract label distributions from logical labels. In this paper, we propose a novel label enhancement method by using adaptive neighbors. Based on the local connectivity, we first assign adaptive neighbors to each instance, then the smoothness constraint assigns similar label distributions to similar instances. Second, we solve the class-imbalance problem by assigning weights to instances. Finally, comparison experiments on 13 datasets show that our method fits the ground-truth label distributions better.