Reproduce “Learning Fair Domain Adaptation with Virtual Label Distribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuguang Zhang, Lijun Sheng, Jian Liang, Ran He
Researchers introduced Virtual Label-distribution-aware Learning (VILL), a framework designed to improve category fairness in Unsupervised Domain Adaptation (UDA). VILL consistently boosted the worst-N accuracy of lowest-performing categories across multiple UDA baselines and datasets while maintaining or slightly improving overall classification performance.
Key insights
- Even with high overall accuracy, existing UDA methods can exhibit significant performance disparities across categories, leading to low worst-case accuracy for certain classes.
- Dynamically estimating target class distributions via pseudo-labels, even without ground truth, can effectively guide re-weighting and re-balancing to improve fairness.
- Improving the performance of underrepresented categories in the target domain requires both re-weighting source supervision and explicitly re-balancing target decision boundaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.