Research map / Learning theory and algorithms / Fairness
Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes II-GOOT, an optimal-transport-based training framework that shares samples among similar disadvantaged users and aligns disadvantaged users with advantaged users to reduce user-oriented unfairness, plus a new evaluation metric ξ-UOF that jointly captures fairness and accuracy.
Problem
Recommender systems perform much better for advantaged (highly active) users than for disadvantaged users (User-Oriented Fairness issue), and existing methods neither fix the unfair training process nor evaluate fairness in a way that distinguishes good-fair from bad-fair outcomes.
Approach
Use intra-group optimal transport (Sinkhorn) to pair similar disadvantaged users for sample sharing, and inter-group optimal clustering to pull disadvantaged user embeddings toward their matched advantaged users; additionally, the ξ-UOF metric measures the gap to the best-UOF state (both groups reaching advantaged-level accuracy), so that uniformly poor fairness is penalized.
Key results
- On LightGCN (Beauty), II-GOOT raises overall NDCG from 0.245 to 0.286 and reduces ξ-UOF from 0.190 to 0.150 versus the original backbone.
- II-GOOT beats UFR and S-DRO on all three datasets and all four backbones (MF, NeuMF, VAECF, LightGCN) in NDCG and HR per Table 1.
- In the ablation on LightGCN, the intra-group stage alone gives a larger gain than the inter-group stage alone (e.g., Beauty NDCG ξ-UOF 0.160 vs 0.159 but Health 0.207 vs 0.219 for intra vs inter).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add II-GOOT to LightGCN | LightGCN · Beauty | NDCG: 0.248 -> 0.286* | Helped |
| add II-GOOT to LightGCN | LightGCN · Grocery | NDCG: 0.260 -> 0.289* | Helped |
| add II-GOOT to LightGCN | LightGCN · Health | NDCG: 0.272 -> 0.309* | Helped |
| add II-GOOT to LightGCN | LightGCN · Beauty | HR: 0.376 -> 0.433* | Helped |
| add II-GOOT to LightGCN | LightGCN · Grocery | HR: 0.414 -> 0.443* | Helped |
| add II-GOOT to LightGCN | LightGCN · Health | HR: 0.387 -> 0.448* | Helped |
| add the intra-group stage (ablation) | LightGCN · Beauty | NDCG: 0.245 -> 0.274 | Helped |
| add the inter-group stage (ablation) | LightGCN · Beauty | NDCG: 0.245 -> 0.269 | Helped |
| add the intra-group stage (ablation) | LightGCN · Grocery | NDCG: 0.257 -> 0.279 | Helped |
| add the inter-group stage (ablation) | LightGCN · Grocery | NDCG: 0.257 -> 0.277 | Helped |
| add the intra-group stage (ablation) | LightGCN · Health | NDCG: 0.263 -> 0.290 | Helped |
| add the inter-group stage (ablation) | LightGCN · Health | NDCG: 0.263 -> 0.280 | Helped |
Surprising findings
- UFR lowers advantaged-user NDCG (0.3046 to 0.1956 on NeuMF Beauty) while disadvantaged NDCG barely rises, showing the existing UOF metric rewards fairness achieved by degrading accuracy.
Limitations
- Advantaged users are defined as the top 5% by interaction count, a fixed threshold whose sensitivity is only tested in an appendix.
- Attaining the ideal Best UOF is acknowledged as infeasible, so ξ-UOF measures only a gap to an unreachable target.
- Intra-group stage is limited by severe data sparsity among disadvantaged users, which constrains its gains.
Experimental setup
Datasets: Amazon Beauty Amazon Grocery & Gourmet Food Amazon Health & Personal Care
Benchmarks: Amazon Beauty Amazon Grocery & Gourmet Food Amazon Health & Personal Care
Metrics: NDCG HR ξ-UOF
Baselines: UFR S-DRO Original backbone
Builds on: UFR S-DRO Sinkhorn distances
Closest papers by meaning
- User-item fairness tradeoffs in recommendations (NeurIPS 2024 · similarity 0.89)
- CUGF: A Reliable and Fair Recommendation Framework (AAAI 2025 · similarity 0.87)
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 2024 · similarity 0.87)
- Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation (ICLR 2025 · similarity 0.87)
- ENSUR: Equitable and Statistically Unbiased Recommendation (ICML 2025 · similarity 0.86)
- Leveraging Opposite Gender Interaction Ratio as a Path towards Fairness in Online Dating Recommendations Based on User Sexual Orientation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.