Research map / Label-efficient and robust learning / Recognition and long-tail
From Pairwise to Ranking: Climbing the Ladder to Ideal Collaborative Filtering with Pseudo-Ranking
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 Pseudo-Ranking Paradigm (PRP) for collaborative filtering: noise-injected pseudo-rankings from a learned ranker, a listwise softplus ranking loss over adjacent sub-rankings, and a gradient-density confidence weighting, improving over pairwise CF baselines on four datasets.
Problem
Pairwise loss functions used in implicit collaborative filtering approximate full user rankings poorly, and real datasets lack full ranking supervision and suitable ranking losses.
Approach
Generate pseudo-rankings of sampled items by injecting Gaussian noise of two magnitudes into interacted items (user-dependent variance via MLPs with reparameterization), train a ranker MLP on them, and use a softplus-based ranking loss over adjacent items in the ranking, weighted by a gradient-density confidence coefficient.
Key results
- On Yelp, PRP reaches HR@10 0.2000 and NDCG@10 0.0464 versus 0.1537 and 0.0334 for BPR, a 30.12% and 38.92% improvement.
- On Gowalla, PRP improves NDCG@20 by 29.35% over BPR (0.1040 vs 0.0804) and outperforms ANS and MixGCF.
- Integrated with MF, PRP improves Yelp Recall@10 from 0.0427 to 0.0577 (about 31.78% average improvement across metrics reported).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PRP instead of BPR | Collaborative filtering · ML-1M | NDCG@10: 10.49% | Helped |
| Use PRP instead of BPR | Collaborative filtering · Gowalla | NDCG@10: 31.16% | Helped |
| Use PRP instead of BPR | Collaborative filtering · Foursquare | NDCG@10: 29.68% | Helped |
| Use PRP instead of BPR | Collaborative filtering · Yelp | NDCG@10: 38.92% | Helped |
| Add PRP to MF | MF · Yelp | NDCG@10: 0.0334 -> 0.0464 | Helped |
| Add PRP to LightGCN | LightGCN · Gowalla | NDCG@10: 0.0917 -> 0.0981 | Helped |
| Add PRP to NGCF | NGCF · Yelp | NDCG@10: 0.0345 -> 0.0406 | Helped |
| Remove the ranker (ablation) | PRP · Figure-reported ablation settings | recommendation performance | Hurt |
| Remove the pseudo-ranking supervision loss L_p (ablation) | PRP · Figure-reported ablation settings | recommendation performance | Hurt |
| Remove the confidence mechanism (ablation) | PRP · Figure-reported ablation settings | recommendation performance | Hurt |
| Increase ranking length beyond 5 (ablation) | PRP ranker · Figure-reported ranking-length analysis | recommendation performance | Hurt |
| Increase supervised magnitude β (ablation) | PRP · Figure-reported supervised-magnitude analysis | recommendation performance | Mixed |
Surprising findings
- Even a simple BPR-MF model with PRP achieves performance comparable to more sophisticated GNN models, according to the authors.
- Performance drops sharply for ranking length k>5, attributed to the ranker's inability to rank accurately.
Limitations
- Performance degrades for large ranking length k>5 due to ranker inaccuracy.
- Pseudo-rankings are inaccurate and rely on noise-injection heuristics without ground-truth ranking validation.
Experimental setup
Datasets: MovieLens-1M Yelp Gowalla Foursquare
Benchmarks: MovieLens-1M Yelp Gowalla Foursquare
Metrics: HR@10 HR@20 Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: BPR SRNS MixGCF UIB SimpleX ANS
Builds on: BPR reparameterization trick RecBole
Closest papers by meaning
- SQL-Rank: A Listwise Approach to Collaborative Ranking (ICML 2018 · similarity 0.90)
- PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation (NeurIPS 2024 · similarity 0.89)
- Learning-Efficient Yet Generalizable Collaborative Filtering for Item Recommendation (ICML 2024 · similarity 0.88)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · similarity 0.86)
- Advancing Loss Functions in Recommender Systems: A Comparative Study with a Rényi Divergence-Based Solution (AAAI 2025 · similarity 0.86)
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · 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.