Heterogeneous domain adaptation remains challenging due to the inconsistent feature representations and data distributions between the source and target domains. In this paper, we introduce contrastive learning into heterogeneous domain adaptation, which is barely considered in the literature. Specifically, we propose a novel model called Prototype Contrastive and Adversarial Alignment (PCAA), taking both feature- and class-level distribution alignments into account to discover an informative shared feature space for the source and target domains. For feature-level distribution matching, we develop an adversarial network to reduce the marginal distribution difference. Besides, PCAA involves a prototype contrastive schema to minimize the conditional distribution discrepancy for class-level distribution alignment. We further reduce the cross-entropy loss on the labeled samples to achieve informative latent representations. Experiments on several real-world datasets validate the effectiveness of the proposed model for image classification.
Sun et al. (Tue,) studied this question.