Research map / Learning theory and algorithms / Fairness
User-item fairness tradeoffs in recommendations
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.
Develops a theoretical model of user-item fairness tradeoffs in recommendation (max-min normalized utility with item-fairness constraints) and validates phenomena on a prototype arXiv preprint recommender.
Problem
Understanding the price of item fairness on user fairness in multi-sided recommendation, including how user preference diversity and preference misestimation (cold-start users) shape the tradeoff.
Approach
Formalizes user fairness as max-min normalized user utility subject to item fairness constraints, proves sparse/symmetric solution structure via a reduction to linear programs, and shows diverse user preferences yield "free" fairness while item constraints can arbitrarily worsen the cost of misestimation for cold-start users.
Key results
- Price of fairness is low for most item-fairness levels (gamma up to 0.9) but rises steeply as gamma approaches 1, and is larger for homogeneous user populations.
- Theory: the price of fairness is decreasing in alpha on (0,1/2] and increasing on [1/2,1), so heterogeneous populations face less cost.
- Theory: with beta > 1/n, item fairness constraints can make the price of misestimation arbitrarily large (>1-epsilon), while it is at most 1/2 without them; empirically, item constraints do not substantially increase misestimation cost on arXiv data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase user preference diversity | Recommendation policy with item fairness constraints · User-item fairness tradeoff | minimum normalized user utility | Helped |
| Apply item fairness constraints | Users with misestimated preferences · Cold-start recommendation, theoretical analysis | minimum normalized user utility | Hurt |
| Apply item fairness constraints | arXiv recommender with misestimated user preferences · arXiv recommendation | minimum normalized user utility | No clear effect |
| Use a homogeneous rather than random user population | arXiv recommender · User-item fairness tradeoff | minimum normalized user utility | Hurt |
| Replace max-min fairness with Nash welfare fairness (ablation) | arXiv recommender · User-item fairness tradeoff | user Nash welfare | Mixed |
| Replace max-min fairness with sum-of-3-min fairness (ablation) | arXiv recommender · User-item fairness tradeoff | sum of worst-off user utilities | Mixed |
| Interpolate item utility from symmetric utility toward exposure utility (ablation) | arXiv recommender · User-item fairness tradeoff | minimum normalized user utility | No clear effect |
| Use max rather than mean similarity (ablation) | TF-IDF arXiv recommender · Future citation prediction | adjusted R-squared: 0.04616 -> 0.08915 | Helped |
Surprising findings
- Empirically, item fairness constraints on average do not increase the cost of misestimation, despite the worst-case theoretical result.
- Under the max-min objective, item-fair solutions can recommend cold-start users the globally least preferred items even when maximizing minimum user utility.
Limitations
- Theory assumes each user receives only a single recommended item.
- Restricted theoretical settings with only two or three user types and opposing preferences.
- Symmetric utility model assumed in main results; alternative item utility models only checked empirically.
Experimental setup
Datasets: arXiv dataset (Kaggle) Semantic Scholar citation data
Benchmarks: Semantic Scholar citation prediction on arXiv CS papers
Metrics: logistic regression coefficient on similarity score minimum normalized user utility minimum normalized item utility price of fairness price of misestimation
Baselines: TF-IDF SPECTER
Builds on: price of fairness two-sided fairness in recommendation exposure-based item fairness
Closest papers by meaning
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 2024 · similarity 0.92)
- Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems (AAAI 2024 · similarity 0.89)
- CUGF: A Reliable and Fair Recommendation Framework (AAAI 2025 · similarity 0.88)
- ENSUR: Equitable and Statistically Unbiased Recommendation (ICML 2025 · similarity 0.87)
- Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation (ICLR 2025 · similarity 0.86)
- FairJob: A Real-World Dataset for Fairness in Online Systems (NeurIPS 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.