TIToK: A solution for bi-imbalanced unsupervised domain adaptation.

Abstract

Unsupervised domain adaptation (UDA) aims to transfer knowledge via domain alignment, and typically assumes balanced data distribution. When deployed in real tasks, however, (i) each domain usually suffers from class imbalance, and (ii) different domains may have different class imbalance ratios. In such bi-imbalanced cases with both within-domain and across-domain imbalance, source knowledge transfer may degenerate the target performance. Some recent efforts have adopted source re-weighting to this issue, in order to align label distributions across domains. However, since target label distribution is unknown, the alignment might be incorrect or even risky. In this paper, we propose an alternative solution named TIToK for bi-imbalanced UDA, by directly Transferring Imbalance-Tolerant Knowledge across domains. In TIToK, a class contrastive loss is presented for classification, in order to alleviate the sensitivity to imbalance in knowledge transfer. Meanwhile, knowledge of class correlation is transferred as a supplementary, which is commonly invariant to imbalance. Finally, discriminative feature alignment is developed for a more robust classifier boundary. Experiments over benchmark datasets show that TIToK achieves competitive performance with the state-of-the-arts, and its performance is less sensitive to imbalance.