Imbalanced Label Distribution Learning

Label distribution covers a certain number of labels, representing the degree to which each label describes an instance. The learning process on instances labeled by distributions is called Distribution Learning (LDL). Although LDL has been applied successfully many practical applications, one problem with existing methods that they are limited data balanced information. However, annotation information in real-world often exhibits imbalanced distributions, significantly degrades performance methods. In this paper, we investigate Imbalanced (ILDL) problem. To handle challenging problem, delve into characteristics ILDL and empirically find representation shift underlying reason for degradation Inspired finding, present novel method named Representation Alignment (RDA). RDA aligns feature representations alleviate impact gap between training set test caused imbalance issue. Extensive experiments verify superior RDA. Our work fills benchmarks techniques problems.