Research map / Learning theory and algorithms / Fairness
ENSUR: Equitable and Statistically Unbiased Recommendation
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 ENSUR, a risk-controlling prediction-set framework for recommender systems that jointly guarantees coverage, user-group fairness, and minimal set size, optimized by a greedy algorithm (GUFA) using concentration-bound upper bounds.
Problem
Recommender systems lack statistical guarantees on recommendation confidence (coverage) and user-group fairness, and minimum-size recommendation sets with such guarantees are not provided by existing methods.
Approach
Adapts risk-controlling prediction sets (RCPS) to recommendation by adding a fairness-controlling constraint on the hit-rate or DCG gap between two user groups, bounds risk with a binomial CDF and fairness with a Bernstein bound, and greedily lowers score thresholds per group to minimize average set size, treating any base recommender as a black box.
Key results
- On AmazonOffice, all ENSUR-enhanced base models keep risk at or below the 0.20 threshold for both groups, while several base models and fairness baselines exceed it for at least one group.
- ENSUR reduces hit rate difference and DCG difference below the 0.20 fairness threshold for all base models across datasets, while producing the smallest average set sizes in most settings.
- Training-time comparison reports ENSUR at 8 to 15 minutes versus 22 to 135 minutes for fairness baselines across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ENSUR to MLP | MLP · AmazonOffice, group 2 | risk: 0.38 -> 0.184 | Helped |
| Apply ENSUR to LightGCN | LightGCN · MovieLens, group 1 | risk: 0.212 † -> 0.128 | Helped |
| Apply ENSUR to MLP | MLP · Last.fm, group 2 | risk: 0.24† -> 0.19 | Helped |
| Apply ENSUR to DeepFM | DeepFM · Book-Crossing, group 2 | risk: 0.429† -> 0.191 | Helped |
| Apply ENSUR to MLP | MLP · AmazonOffice | Hit Rate Diff: 0.219† -> 0.013 | Helped |
| Apply ENSUR to MLP | MLP · MovieLens | DCG Diff: 0.007 -> 0.004 | Helped |
| Apply ENSUR to NeuMF | NeuMF · Last.fm | DCG Diff: 0.021 -> 0.004 | Helped |
| Apply ENSUR to DeepFM | DeepFM · Book-Crossing | Hit Rate Diff: 0.302† -> 0.003 | Helped |
| Increase the risk threshold α (ablation) | ENSUR · AmazonOffice | average prediction set size | Helped |
| Increase risk confidence δ (ablation) | ENSUR · Book-Crossing | average prediction set size | Helped |
| Increase fairness threshold η (ablation) | ENSUR · MovieLens | average prediction set size | Helped |
| Increase fairness confidence ˆδ (ablation) | ENSUR · Last.fm | average prediction set size | Helped |
Limitations
- Upper bounds on risk and fairness may be loose, which affects optimization efficiency (acknowledged by authors).
- Risk and fairness thresholds are chosen by manual validation, and the fairness analysis covers only two groups.
- Theoretical guarantees rely on independence assumptions across users and groups, which the paper itself assumes (Assumption A.1).
Experimental setup
Datasets: AmazonOffice Last.fm MovieLens Book-Crossing
Benchmarks: AmazonOffice MovieLens Last.fm Book-Crossing
Metrics: risk average set size hit rate difference DCG difference
Baselines: NFCF MFCF GMF-UFR NeuMF-UFR
Builds on: Risk-Controlling Prediction Sets (RCPS) Bernstein inequality UFR (user fairness re-ranking)
Closest papers by meaning
- CUGF: A Reliable and Fair Recommendation Framework (AAAI 2025 · similarity 0.88)
- User-item fairness tradeoffs in recommendations (NeurIPS 2024 · similarity 0.87)
- Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems (AAAI 2024 · similarity 0.86)
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 2024 · similarity 0.86)
- Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation (ICLR 2025 · similarity 0.85)
- Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommender Systems (NeurIPS 2026 · similarity 0.84)
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.