Research map / Learning theory and algorithms / Fairness
Leveraging Opposite Gender Interaction Ratio as a Path towards Fairness in Online Dating Recommendations Based on User Sexual Orientation
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 Opposite Gender Interaction Ratio (OGIR) to measure sexual-orientation-related unfairness in online dating recommenders, finds group unfairness with MF/NGCF/CAGCN*, and mitigates it via re-weighting (group quantity) and calibration-based re-ranking (MMR).
Problem
Recommender systems in online dating may give unequal recommendation quality to users with different sexual orientations (proxied by preference for the opposite gender), a fairness dimension that has been under-studied.
Approach
An interaction-based metric (OGIR, the fraction of a user's rated users who are of the opposite gender) groups users without explicit orientation labels; group unfairness is measured as the normalized average pairwise utility gap, and two fixes target group quantity imbalance (group-size re-weighting in BPR loss) and calibration imbalance (MMR-based gender-ratio re-ranking).
Key results
- Re-weighting improved average fairness by 18.33% (MF), 24.44% (NGCF), and 5.04% (CAGCN*) with utility drops of at most about 4%.
- Re-ranking with MMR improved both utility and fairness for baseline models, avoiding the usual fairness-utility tradeoff, and the combined re-weighting plus re-ranking model achieved the best fairness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Re-weight BPR training by OGIR-group size | MF · Líbímseti.cz test | Avg Fairness: 0.5449 -> 0.4450 (+18.33%) | Helped |
| Re-weight BPR training by OGIR-group size | MF · Líbímseti.cz test | Avg Utility: 0.2023 -> 0.2019 (-0.20%) | Hurt |
| Re-weight BPR training by OGIR-group size | NGCF · Líbímseti.cz test | Avg Fairness: 0.5692 -> 0.4301 (+24.44%) | Helped |
| Re-weight BPR training by OGIR-group size | NGCF · Líbímseti.cz test | Avg Utility: 0.2043 -> 0.1960 (-4.06%) | Hurt |
| Re-weight BPR training by OGIR-group size | CAGCN* · Líbímseti.cz test | Avg Fairness: 0.5635 -> 0.5351 (+5.04%) | Helped |
| Re-weight BPR training by OGIR-group size | CAGCN* · Líbímseti.cz test | Avg Utility: 0.2267 -> 0.2244 (-1.01%) | Hurt |
| Add calibration-aware re-ranking to baseline recommendations | MF, NGCF, and CAGCN* · Líbímseti.cz test | Avg Fairness | Helped |
| Add calibration-aware re-ranking to baseline recommendations | MF, NGCF, and CAGCN* · Líbímseti.cz test | Avg Utility | Helped |
| Add re-ranking after re-weighting | MF, NGCF, and CAGCN* · Líbímseti.cz test | Avg Fairness | Helped |
| Increase re-weighting strength p (ablation) | Re-weighted recommenders · Líbímseti.cz test | Utility | Hurt |
| Increase re-weighting strength p (ablation) | Re-weighted recommenders · Líbímseti.cz test | Fairness | Helped |
| Increase re-ranking trade-off parameter λ (ablation) | MF re-ranking · Líbímseti.cz test | Utility | Helped |
Surprising findings
- Re-ranking improved utility as well as fairness, so the usual fairness-utility tradeoff did not appear.
- Simply selecting models by a utility-minus-fairness score and training separate models per group were judged insufficient fixes.
Limitations
- Only one real-world dating dataset (Líbímseti.cz) was used, and only binary gender is available.
- OGIR is an implicit interaction-based proxy rather than an explicit sexual orientation label.
- Re-ranking relies on gender features, and the authors' test with gender-injected embeddings did not show that gender information is the key driver.
Experimental setup
Datasets: Líbímseti.cz deezer-europe
Benchmarks: Líbímseti.cz deezer-europe
Metrics: Recall@20 Precision@20 F1@20 Hit Ratio@20 NDCG@20 group unfairness Delta metrics
Baselines: MF NGCF CAGCN* Model re-weighting only Model re-ranking only
Builds on: BPR Maximum Marginal Relevance NGCF LightGCN CAGCN
Closest papers by meaning
- Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems (AAAI 2024 · similarity 0.85)
- User-item fairness tradeoffs in recommendations (NeurIPS 2024 · similarity 0.84)
- FairJob: A Real-World Dataset for Fairness in Online Systems (NeurIPS 2024 · similarity 0.84)
- No Prejudice! Fair Federated Graph Neural Networks for Personalized Recommendation (AAAI 2024 · similarity 0.83)
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 2024 · similarity 0.83)
- CUGF: A Reliable and Fair Recommendation Framework (AAAI 2025 · similarity 0.82)
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.