Maintaining Structural Information by Pairwise Similarity for Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) aims to transfer the knowledge learned from the labeled source domain to the unlabeled target domain. Among them, the source domain and the target domain have the same label space, but the representation distributions of their input space are different. Mainstream approaches resort to domain adversarial training to align input distributions of two domains in the feature space. Although these methods have made remarkable progress, they have the risk of destroying discriminative structural information between different classes in the target domain. To alleviate this risk, we are inspired by the problem reduction method in ensemble methods and binarization techniques, and propose a novel approach Maintaining Structural Information of the target domain based on Pairwise semantic Similarity (whether two instances belong to the same class or not) (MSIPS). Specifically, We introduce Contrastive Learning to obtain feature prototypes for each category on the source domain, and then use these prototypes to predict the similarity of paired target domain samples. Finally, we restrict the target domain to maintain discriminative structural information through such weak information (i.e., pairwise similarity). Extensive experiments of various domain shift scenarios show that our method obtains competitive performance with SOTA, and qualitative visualization can demonstrate the effectiveness of our method.