Subdomain Adaptation Method Based on Transferable Semantic Alignment and Class Correlation

Qian Han^1*Jinfu Lao²Jinyong Zhang¹

• ¹Department of Ideological and Political Theory Teaching, Maoming Polytechnic, Maoming, Guangdong 525000, China, Maoming, China
• ²China Mobile Group Design Institute Co Ltd, Beijing, China

To address the challenges in deep unsupervised domain adaptation within transfer learning—specifically the inadequate utilization of semantic relationships among samples and the excessive complexity of pseudo-label optimization—this study proposes a subdomain adaptation method driven by transferable semantic alignment and class correlation. First, source and target domains are partitioned into subdomains based on class labels, and a novel joint subdomain distribution alignment mechanism is introduced. This mechanism aims to reduce intra-class distribution divergence across domains while enlarging inter-class disparities, thereby effectively mitigating domain shift. Second, a domain-adaptive semantic consistency loss is introduced to enhance the clustering of samples with similar semantics while maximizing the distinction between those with differing semantic meanings in the unified representation space, enabling precise semantic alignment between domains. Third, to improve the quality of pseudo-labels in the target domain, a temperature-based label smoothing technique is adopted. In addition, a class correlation matrix is constructed. Furthermore, a distinctive loss function capturing inter-class relationships is designed to mine intrinsic inter-class and intra-class relationships. This enables the framework to capture within-category coherence and between-category distinction without introducing extra architectural components, thereby enhancing recognition performance on the target domain. Extensive experiments on multiple public datasets demonstrate the superior average classification accuracy of the proposed method. This article's code is open source in https://github.com/XXXX/XXXXX.