CentroIDA: Cross-domain class discrepancy minimization based on accumulative class-centroids for Imbalanced Domain Adaptation

Abstract

Unsupervised Domain Adaptation (UDA) is a critical task in transfer learning, which seeks to apply the knowledge acquired in one domain (the source domain) to a different but related domain (the target domain). In this context, the two domains exhibit Out-Of-Distribution(OOD), and the label information of the target domain is unavailable. Although numerous UDA methods have made significant strides in mitigating the discrepancies in feature distributions across domains, they often neglect a more prevalent real-world scenario. This scenario, referred to as Imbalanced Domain Adaptation (IDA) task, is characterized by the coexistence of differences in both feature and label distributions. Our empirical analysis suggests that overlooking the disparities in label distributions can severely compromise the performance of the model, thereby limiting the generalizability of domain adaptation models. Consequently, our objective is to introduce a model capable of concurrently addressing differences in both feature and label distributions, thereby fostering a broader application of domain adaptation methodologies. In this paper, we propose a cross-domain class discrepancy minimization method based on accumulative class-centroids for IDA, termed as centroIDA. The strategies of accumulative class-centroids alignment and class-wise feature alignment pay heed to the relationship between cross-domain intra-class features and inter-class features, thereby facilitating conditional feature alignment to overcome both feature and label distribution differences simultaneously. Our experimental evaluation on OfficeHome and DomainNet datasets reveals that, in comparison to TIToK (a novel method also designed to tackle the IDA problem), our method enhances the average accuracy by 5.6% and 2.7%, respectively. Furthermore, as the label shift intensifies, our method consistently outperforms other methods, demonstrating a robust resistance to label shift.