Prototype Combination for Multi-Source Unsupervised Domain Adaptation

Multi-source unsupervised domain adaptation (MSUDA) is a technique that transfers knowledge from multiple labeled source domains to an unlabeled target domain. The challenge of MSUDA is to reduce the domain shift and effectively amalgamate knowledge from disparate source domains. To address this challenge, it is necessary to model the target domain as a weighted combination of the source domains at the category level. Therefore, we propose a prototype combination method for multi-source unsupervised domain adaptation, which establishes multiple domain alignment in a combinatorial manner. Our method is established on a set of semantic category prototypes, each of which is a representative category embedding. A prototype combination mechanism (i.e., a feature-fusion scheme) is designed to select which source class features should be aligned with the corresponding target class features. This method incorporates contrastive prototype adaptation (i.e., a category-wise alignment approach) to accommodate the label distributions of the target domain. Furthermore, a prototype combination regularization (i.e., a domain-wise alignment metric) is designed to reduce the distributional differences between the source category prototypes and the target samples of low-quality pseudo-labels. The experimental results on three benchmark datasets demonstrate that our prototype combination mechanism is capable of selecting and combining category-discriminative features across multiple source domains, while the prototype combination regularization can further reduce the domain shift.