Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Double-Edged Sword of Knowledge Transfer: Diagnosing and Curing Fairness Pathologies in Cross-Domain Recommendation

Yuhan Zhao, Weixin Chen, Li Chen, Weike Pan

TL;DR

This work analyzes why cross-domain recommendation can introduce group unfairness by identifying two mechanisms: Cross-Domain Disparity Transfer and unfairness from Cross-Domain Information Gain. It proposes Cross-Domain Fairness Augmentation (CDFA), a model-agnostic framework with two modules: unlabeled-data augmentation to balance informative signals across user groups, and an information-theoretic gain redistribution to equalize cross-domain benefits, both underpinned by a theoretical bound framework using $W_1$ (Wasserstein-1) and uniform-convergence analyses. Empirically, CDFA improves fairness metrics while preserving or boosting accuracy across multiple CDR backbones and datasets, outperforming existing fairness-aware CDR methods like FairCDR and VUG. The results demonstrate a practical path to fairer, more trustworthy cross-domain recommendations with broad applicability in real-world systems, supported by rigorous theory and extensive experiments.

Abstract

Cross-domain recommendation (CDR) offers an effective strategy for improving recommendation quality in a target domain by leveraging auxiliary signals from source domains. Nonetheless, emerging evidence shows that CDR can inadvertently heighten group-level unfairness. In this work, we conduct a comprehensive theoretical and empirical analysis to uncover why these fairness issues arise. Specifically, we identify two key challenges: (i) Cross-Domain Disparity Transfer, wherein existing group-level disparities in the source domain are systematically propagated to the target domain; and (ii) Unfairness from Cross-Domain Information Gain, where the benefits derived from cross-domain knowledge are unevenly allocated among distinct groups. To address these two challenges, we propose a Cross-Domain Fairness Augmentation (CDFA) framework composed of two key components. Firstly, it mitigates cross-domain disparity transfer by adaptively integrating unlabeled data to equilibrate the informativeness of training signals across groups. Secondly, it redistributes cross-domain information gains via an information-theoretic approach to ensure equitable benefit allocation across groups. Extensive experiments on multiple datasets and baselines demonstrate that our framework significantly reduces unfairness in CDR without sacrificing overall recommendation performance, while even enhancing it.

The Double-Edged Sword of Knowledge Transfer: Diagnosing and Curing Fairness Pathologies in Cross-Domain Recommendation

TL;DR

This work analyzes why cross-domain recommendation can introduce group unfairness by identifying two mechanisms: Cross-Domain Disparity Transfer and unfairness from Cross-Domain Information Gain. It proposes Cross-Domain Fairness Augmentation (CDFA), a model-agnostic framework with two modules: unlabeled-data augmentation to balance informative signals across user groups, and an information-theoretic gain redistribution to equalize cross-domain benefits, both underpinned by a theoretical bound framework using (Wasserstein-1) and uniform-convergence analyses. Empirically, CDFA improves fairness metrics while preserving or boosting accuracy across multiple CDR backbones and datasets, outperforming existing fairness-aware CDR methods like FairCDR and VUG. The results demonstrate a practical path to fairer, more trustworthy cross-domain recommendations with broad applicability in real-world systems, supported by rigorous theory and extensive experiments.

Abstract

Cross-domain recommendation (CDR) offers an effective strategy for improving recommendation quality in a target domain by leveraging auxiliary signals from source domains. Nonetheless, emerging evidence shows that CDR can inadvertently heighten group-level unfairness. In this work, we conduct a comprehensive theoretical and empirical analysis to uncover why these fairness issues arise. Specifically, we identify two key challenges: (i) Cross-Domain Disparity Transfer, wherein existing group-level disparities in the source domain are systematically propagated to the target domain; and (ii) Unfairness from Cross-Domain Information Gain, where the benefits derived from cross-domain knowledge are unevenly allocated among distinct groups. To address these two challenges, we propose a Cross-Domain Fairness Augmentation (CDFA) framework composed of two key components. Firstly, it mitigates cross-domain disparity transfer by adaptively integrating unlabeled data to equilibrate the informativeness of training signals across groups. Secondly, it redistributes cross-domain information gains via an information-theoretic approach to ensure equitable benefit allocation across groups. Extensive experiments on multiple datasets and baselines demonstrate that our framework significantly reduces unfairness in CDR without sacrificing overall recommendation performance, while even enhancing it.
Paper Structure (33 sections, 3 theorems, 30 equations, 4 figures, 3 tables)

This paper contains 33 sections, 3 theorems, 30 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Theoretical Analysis of Fairness Violations in CDR
  3. Notation
  4. Cross-Domain Disparity Transfer
  5. Unfairness from Cross-Domain Information Gain
  6. Sufficient Condition for Fairness Preservation in CDR
  7. Empirical Evaluation
  8. Methodology
  9. Cross-Domain Disparity Transfer
  10. Cross-Domain Gain Redistribution
  11. Training Procedure
  12. EXPERIMENTS
  13. Experimental Setup
  14. Overall Performance Comparison
  15. Performance Comparison with Other Fairness-aware CDR Methods
  16. ...and 18 more sections

Key Result

theorem 1

Assume $f_t: (\mathcal{Z}, d_\mathcal{Z}) \to (\mathcal{Y}, d_\mathcal{Y})$ is $L_f$-Lipschitz and $o: (\mathcal{Y}, d_\mathcal{Y}) \to \mathbb{R}$ is $L_o$-Lipschitz and bounded by $B$. Then the target-domain group fairness satisfies Moreover, by the triangle inequality, and with the abbreviations we have the bound Consequently,

Figures (4)

  • Figure 1: UGF comparison across settings on the Tenrec QB dataset, for the sensitive attribute gender.
  • Figure 2: The illustration of our proposed Cross-Domain Fairness Augmentation (CDFA) framework.
  • Figure 3: Comparison of fairness-aware cross-domain recommendation baselines on the Tenrec QB and QK datasets.
  • Figure 4: Impact of the hyperparameters on our CDFA on the Tenrec QB dataset.

Theorems & Definitions (14)

  • Definition 2.1.1: User-Oriented Group Fairness (UGF) Measure
  • Definition 2.1.2: Cross-Domain Recommendation System
  • theorem 1: Upper Bound under Domain Shift and Group Imbalance
  • proof : Proof sketch
  • Remark 1
  • theorem 2: Uniform Convergence for Group-wise Gain Gaps
  • proof : Proof sketch
  • Remark 2
  • theorem 3: Fairness Preservation via Upper Bounds
  • proof : Proof sketch
  • ...and 4 more