Research map / Learning theory and algorithms / Fairness
Interpolating Item and User Fairness in Multi-Sided 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.
Proposes a constrained multi-sided fair recommendation framework (platform revenue maximization with item and user fairness constraints) and FORM, a low-regret online algorithm combining relaxed constraints with randomized exploration, evaluated on Amazon clothing reviews.
Problem
Recommendation on multi-sided platforms must balance platform revenue with fairness for items (sellers) and users, but online learning under unknown purchase/arrival probabilities makes fairness constraints impossible to verify exactly.
Approach
Formulates fair recommendation as a constrained optimization that interpolates item and user fairness via parameters, then solves a relaxed version on estimated instances with randomized exploration so learning and fairness proceed simultaneously with sublinear regret.
Key results
- Revenue regret and fairness regret of FORM are bounded by O(M N^{1/3} T^{-1/3}) under local Lipschitz assumption.
- FairRec and TFROM reduce platform revenue by about 38%, while FORM with delta^I=delta^U=0.2 keeps revenue loss within 10%.
- On Amazon data, FORM keeps all items near or above the item-fairness level for both high (0.8) and moderate (0.2) delta^I.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FORM with δI and δU set to 0.2 | Amazon review data recommendation · Amazon review data | revenue loss: within 10% | Helped |
| Use FairRec instead of a revenue-focused recommendation | Amazon review data recommendation · Amazon review data | revenue: about 38% | Hurt |
| Use TFROM instead of a revenue-focused recommendation | Amazon review data recommendation · Amazon review data | revenue: about 38% | Hurt |
| Increase the item fairness parameter δI | Fair recommendation framework · Amazon review data | price of fairness | Hurt |
| Increase the user fairness parameter δU | Fair recommendation framework · Amazon review data | price of fairness | Mixed |
| Add randomized exploration to recommendations | FORM · Online fair recommendation | revenue regret: O(MN^{\frac{1}{3}}T^{-\frac{1}{3}}) | Helped |
| Relax fairness constraints using the estimated instance | FORM · Online fair recommendation | fairness regret: O(MN^{\frac{1}{3}}T^{-\frac{1}{3}}) | Helped |
| Prioritize platform revenue with the greedy baseline | Amazon review data recommendation · Amazon review data | revenue | Helped |
| Use max-utility recommendations | Amazon review data recommendation · Amazon review data | item fairness | Hurt |
| Use min-revenue recommendations | Amazon review data recommendation · Amazon review data | item fairness | Helped |
| Set item and user fairness parameters to 0.8 for the relevant fairness constraint | FORM · Amazon review data | item outcomes relative to item-fair outcomes: δI = 0.8 | Helped |
Surprising findings
- Item fairness is much harder to achieve than user fairness in the Amazon case study, since user MNL utility aligns well with platform revenue.
Limitations
- Assumes known outcome functions and local Lipschitz continuity of the solutions; theory requires bounded-away-from-boundary instances.
- Stochastic i.i.d. arrivals in the core model; adversarial environments left to future work.
- Case study on a single Amazon dataset with simulated arrivals and fixed N=30 items.
Experimental setup
Datasets: Amazon review dataset (Clothing, Shoes and Jewelry)
Benchmarks: Amazon review dataset (Clothing, Shoes and Jewelry) case study
Metrics: platform revenue normalized by OPT-rev revenue regret fairness regret item outcome normalized by item-fair solution user outcome normalized by user-fair solution price of fairness
Baselines: greedy max-utility min-revenue random FairRec TFROM
Builds on: FairRec online learning with knapsack Hoffman bound
Closest papers by meaning
- User-item fairness tradeoffs in recommendations (NeurIPS 2024 · similarity 0.92)
- CUGF: A Reliable and Fair Recommendation Framework (AAAI 2025 · similarity 0.87)
- Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems (AAAI 2024 · similarity 0.87)
- FairJob: A Real-World Dataset for Fairness in Online Systems (NeurIPS 2024 · similarity 0.87)
- Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation (ICLR 2025 · similarity 0.86)
- ENSUR: Equitable and Statistically Unbiased Recommendation (ICML 2025 · similarity 0.86)
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.