Research map / Learning theory and algorithms / Fairness
CUGF: A Reliable and Fair Recommendation Framework
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 CUGF, a conformal-prediction-based recommendation framework that builds user-group-fair prediction sets with coverage guarantees on top of black-box recommenders (DeepFM, LightGCN, GMF, MLP, NeuMF).
Problem
Recommender systems lack statistically guaranteed credibility (uncertainty quantification) and often treat advantaged and disadvantaged user groups unequally; existing fixes address only one issue or are dataset-specific.
Approach
Combines conformal prediction with group fairness: a group-specific threshold parameter gamma defines nested prediction sets, with coverage guaranteed at 1-alpha and a bound eta on the hit-rate or NDCG risk gap between groups, optimized by a greedy CUGF algorithm.
Key results
- With alpha=0.20 and eta=0.20, all five base models met both coverage and fairness thresholds on all four dataset variants.
- Average set sizes ranged roughly from 31 to 50 items across models and datasets; LightGCN had the smallest set on AmazonOffice (36) and MovieLens (36) under the interaction grouping.
- Increasing alpha reduced average set size and increased Diff. (HR), while increasing eta reduced set size in a cascading manner.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CUGF to recommendation models | DeepFM, LightGCN, GMF, MLP, and NeuMF · AmazonOffice and MovieLens, both groupings | coverage and fairness guarantees | Helped |
| Increase the coverage parameter α | CUGF recommendation models · AmazonOffice, both groupings | average set size | Helped |
| Increase the fairness threshold η | CUGF recommendation models · MovieLens, both groupings | average set size | Helped |
| Increase the coverage parameter α | CUGF recommendation models · AmazonOffice, both groupings | Diff. (HR) | Hurt |
| Increase the fairness threshold η | CUGF recommendation models · MovieLens, interactions and popular-items grouping | Diff. (NDCG) | Mixed |
| Use LightGCN as the base model rather than the other tested models | CUGF · AmazonOffice and MovieLens, interactions grouping | average set size: 36 | Helped |
| Use MLP as the base model rather than the other tested models | CUGF · AmazonOffice, interactions and popular-items grouping | average set size: 31 | Helped |
| Use NeuMF as the base model rather than the other tested models | CUGF · MovieLens, interactions and popular-items grouping | average set size: 35 | Helped |
Surprising findings
- The best-performing model for fairness changes across datasets and metrics, so no single base model dominates fairness.
Limitations
- Relies on negative sampling of 50 non-interacted items per user, so evaluation is on a sampled candidate set.
- Set sizes are large (31-50 items), limiting practical precision of recommendations.
- Alpha and eta chosen by manual validation and a bottom-line heuristic rather than principled selection.
Experimental setup
Datasets: AmazonOffice MovieLens
Benchmarks: AmazonOffice (grouped by interactions; interactions + popular items) MovieLens (grouped by interactions; interactions + popular items)
Metrics: average set size Diff. (HR) Diff. (NDCG) coverage
Baselines: DeepFM LightGCN GMF MLP NeuMF
Builds on: conformal prediction adaptive prediction sets (Romano et al. 2020) user-oriented fairness (Li et al. 2021) Neural Collaborative Filtering
Closest papers by meaning
- ENSUR: Equitable and Statistically Unbiased Recommendation (ICML 2025 · similarity 0.88)
- User-item fairness tradeoffs in recommendations (NeurIPS 2024 · similarity 0.88)
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 2024 · similarity 0.87)
- Intra- and Inter-group Optimal Transport for User-Oriented Fairness in Recommender Systems (AAAI 2024 · similarity 0.87)
- FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems (ICML 2025 · similarity 0.85)
- Bridging Jensen Gap for Max-Min Group Fairness Optimization in Recommendation (ICLR 2025 · 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.