Robust domain adaptation with noisy and shifted label distribution

Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain where the data distribution differs. While previous UDA methods have been successful when source domain labels are clean, obtaining clean labels itself is costly. In the presence of label noise in the source domain, traditional UDA methods suffer degradation as they do not handle label noise. This problem has garnered significant attention, leading to the exploration of robust UDA techniques. However, existing robust UDA methods often filter out noisy samples in the source domain or focus on minimizing distribution discrepancies, neglecting the recycling of low-quality noisy samples and ignoring label distribution shift (LDS) issues .

To solve the problems, a research team led by Shao-Yuan LI published their new research on 15 Mar 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

The team proposed an approach named Robust Self-training with Label Refinement (RSLR) to address the above issue. RSLR adopts the self-training framework by maintaining a Labeling Network (LNet) on the source domain, which is used to provide confident pseudo-labels to target samples, and a Target-specific Network (TNet) trained by using the pseudo-labeled samples. To combat the effect of label noise, LNet progressively distinguishes and refines the mislabeled source samples. In combination with class re-balancing to combat the label distribution shift issue, RSLR achieves effective performance on extensive benchmark datasets.

Future work can focus on enhancing the robustness of RSLR against label noise and data distribution shifts, exploring its applicability across different domains, improving model interpretability, integrating with other deep learning techniques, and validating its effectiveness in real-world applications.

DOI: 10.1007/s11704-024-3810-0