Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Fair Domain Adaptation with Virtual Label Distribution

Yuguang Zhang, Lijun Sheng, Jian Liang, Ran He

TL;DR

The paper addresses category fairness in unsupervised domain adaptation by introducing Worst-N accuracy as a metric and proposing VILL, a plug-and-play framework with a virtual label distribution–based re-weighting and a KL-divergence–based re-balancing objective to boost underrepresented classes without harming overall performance. The approach dynamically updates target-derived weights and explicitly refines decision boundaries for minority categories, yielding improved worst-case accuracy across multiple UDA baselines. Key contributions include formalizing worst-case fairness in UDA, a generic VILL framework with two complementary modules, and extensive experiments on OfficeHome and Office showing consistent fairness gains and robust performance. This work offers practical implications for creating more reliable transfer learning systems, particularly in safety- and bias-sensitive applications.

Abstract

Unsupervised Domain Adaptation (UDA) aims to mitigate performance degradation when training and testing data are sampled from different distributions. While significant progress has been made in enhancing overall accuracy, most existing methods overlook performance disparities across categories-an issue we refer to as category fairness. Our empirical analysis reveals that UDA classifiers tend to favor certain easy categories while neglecting difficult ones. To address this, we propose Virtual Label-distribution-aware Learning (VILL), a simple yet effective framework designed to improve worst-case performance while preserving high overall accuracy. The core of VILL is an adaptive re-weighting strategy that amplifies the influence of hard-to-classify categories. Furthermore, we introduce a KL-divergence-based re-balancing strategy, which explicitly adjusts decision boundaries to enhance category fairness. Experiments on commonly used datasets demonstrate that VILL can be seamlessly integrated as a plug-and-play module into existing UDA methods, significantly improving category fairness.

Learning Fair Domain Adaptation with Virtual Label Distribution

TL;DR

The paper addresses category fairness in unsupervised domain adaptation by introducing Worst-N accuracy as a metric and proposing VILL, a plug-and-play framework with a virtual label distribution–based re-weighting and a KL-divergence–based re-balancing objective to boost underrepresented classes without harming overall performance. The approach dynamically updates target-derived weights and explicitly refines decision boundaries for minority categories, yielding improved worst-case accuracy across multiple UDA baselines. Key contributions include formalizing worst-case fairness in UDA, a generic VILL framework with two complementary modules, and extensive experiments on OfficeHome and Office showing consistent fairness gains and robust performance. This work offers practical implications for creating more reliable transfer learning systems, particularly in safety- and bias-sensitive applications.

Abstract

Unsupervised Domain Adaptation (UDA) aims to mitigate performance degradation when training and testing data are sampled from different distributions. While significant progress has been made in enhancing overall accuracy, most existing methods overlook performance disparities across categories-an issue we refer to as category fairness. Our empirical analysis reveals that UDA classifiers tend to favor certain easy categories while neglecting difficult ones. To address this, we propose Virtual Label-distribution-aware Learning (VILL), a simple yet effective framework designed to improve worst-case performance while preserving high overall accuracy. The core of VILL is an adaptive re-weighting strategy that amplifies the influence of hard-to-classify categories. Furthermore, we introduce a KL-divergence-based re-balancing strategy, which explicitly adjusts decision boundaries to enhance category fairness. Experiments on commonly used datasets demonstrate that VILL can be seamlessly integrated as a plug-and-play module into existing UDA methods, significantly improving category fairness.
Paper Structure (12 sections, 8 equations, 2 figures, 2 tables)

This paper contains 12 sections, 8 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Preliminaries
  4. Re-weighting Mechanism for Fair DA
  5. Re-balancing Strategy for Fair DA
  6. Overall Objective
  7. Experiment
  8. Comparison Results
  9. Ablation study
  10. Conclusion
  11. Acknowledgements
  12. Ethical Standards Statement

Figures (2)

  • Figure 1: Results of WAcc-5, WAcc-10 and conventional accuracy of CDAN long2018conditional on 4 tasks from OfficeHome dataset. Worst-$N$ is defined as the average accuracy of $N$ classes with the worst accuracy in the target domain.
  • Figure 2: Worst-N accuracy and global accuracy of VILL on CDAN and PDA on Office dataset.