Causal view mechanism for adversarial domain adaptation

• Published in: Multimedia tools and applications Vol. 82; no. 30; pp. 47347 - 47366
• Main Authors: Fu, Zihao, Wang, Shengsheng, Zhao, Xin, Long, Sifan, Wang, Bilin
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2023; Springer Nature B.V
• Subjects: Adaptation; Causality; Classification; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Datasets; Deep learning; Error reduction; Invariants; Multimedia; Multimedia Information Systems; Representations; Special Purpose and Application-Based Systems; Training; Deep learning; Causal inference; Domain adaptation; Transfer learning; Image classification
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-15683-5

Studies show that the challenge for adversarial domain adaptation is learning domain-invariant representations and alleviating the domain gap. However, the construction of domain-invariant representations suppresses the class-level structure information, and the pursuit of class-level structure information distorts the constructed domain-invariant representations. Till present, it still has difficulty to capture the domain-invariant representations while preserving the class-level structure information in the adversarial training process explicitly. In this paper, we propose a Causal view Mechanism for adversarial Domain Adaptation (CMDA). Firstly, a causal effect model of adversarial DA is proposed and reveals the influence of potential confounders in the adversarial training process. Then, CMDA is proposed to disentangle the domain-specific representations into multiple underlying factors and filter out irrelevant confounding characteristics. Specifically, CMDA capture the desired domain-invariant representations by subtracting the domain-level and class-level confounding characteristics. CMDA not only could preserve the class-level structure information to reduce classification error, but also improve transferability simultaneously. Finally, experiments carried out on Office-31, Office-Home, and VisDA-2017 datasets show that our CMDA method presents strong competition among some recent domain adaptation methods, and the average accuracies achieve 71.3%, 89.3% and 76.1% respectively.