Research map / Label-efficient and robust learning / Recognition and long-tail
Unbiased Recommender Learning from Implicit Feedback via Weakly Supervised Learning
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.
Reframes implicit-feedback recommendation as positive-unlabeled weakly supervised learning (WeaklyRec) and estimates the class prior with progressive proximal transport (PPT), removing the need for negative samples or propensity scores.
Problem
Implicit feedback recommendation suffers from missing negative feedback: unlabeled interactions are treated as negatives, misclassifying hidden positives, and existing unbiased estimators rely on unreliable propensity scores and can be non-negative-violating.
Approach
Derive a PU-learning surrogate risk for recommendation that uses only positive and unlabeled samples, and estimate the unknown class prior by a proximal optimal transport problem whose mass weight is progressively tuned so the transport cost reveals the positive fraction (kappa_p = 1/w*).
Key results
- WeaklyRec reaches NDCG@1 of 0.784 on Yahoo! R3 versus 0.762 for the best baseline DCF (mean over 5 runs).
- On KuaiRec, WeaklyRec achieves NDCG@5 of 0.488 and Recall@5 of 0.296, with significance on most metrics.
- PPT's estimated class prior matches the best-performing fixed prior on KuaiRec (about 0.05), and on a simulated dataset with true prior 0.3 it estimates between 0.25 and 0.33.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WeaklyRec instead of the strongest baseline | Implicit feedback recommendation · Yahoo! R3 | NDCG@5: 0.829±0.003 -> 0.843∗ ±0.002 | Helped |
| Use WeaklyRec instead of the strongest baseline | Implicit feedback recommendation · Coat | Recall@5: 0.389±0.031 -> 0.425∗ ±0.019 | Helped |
| Use WeaklyRec instead of the strongest baseline | Implicit feedback recommendation · KuaiRec | NDCG@5: 0.466±0.004 -> 0.488∗ ±0.004 | Helped |
| Use WeaklyRec instead of the strongest baseline | Implicit feedback recommendation · KuaiRec | Recall@5: 0.288±0.004 -> 0.296∗ ±0.004 | Helped |
| Replace the WeaklyRec risk estimator with uPU, nnPU, or distPU (ablation) | WeaklyRec on KuaiRec · KuaiRec | Recommendation performance (NDCG@K and Recall@K) | Hurt |
| Replace the logistic error measure with the sigmoid error measure (ablation) | WeaklyRec on KuaiRec · KuaiRec | Recommendation performance (NDCG@K and Recall@K) | Hurt |
| Replace the logistic error measure with the squared error measure (ablation) | WeaklyRec on KuaiRec · KuaiRec | Recommendation performance (NDCG@K and Recall@K) | No clear effect |
| Set the fixed class prior to ˆκp = 0.05 instead of ˆκp = 0.2 (ablation) | WeaklyRec · KuaiRec | NDCG@5: 0.4232 -> 0.4871 | Helped |
| Increase the proportion of unlabeled training samples from 5% to all unlabeled samples (ablation) | WeaklyRec · KuaiRec | NDCG@1: approximately 0.34 -> around 0.5 | Helped |
Surprising findings
- Pointwise estimator choice matters little: uPU, nnPU and distPU all perform reasonably within the framework, with uPU overfitting due to unbounded loss.
Limitations
- Evaluated only on ID-based matrix factorization without user profile or content features.
- Unbiasedness depends on accurate class prior estimation; the non-negative clipping of per-term losses breaks exact unbiasedness.
- Theoretical guarantees are stated as consistency without full proof of performance gains over all baselines.
Experimental setup
Datasets: Yahoo! R3 Coat KuaiRec
Benchmarks: Yahoo! R3 Coat KuaiRec
Metrics: NDCG@1 NDCG@3 NDCG@5 Recall@1 Recall@3 Recall@5
Baselines: RMF CRMF WMF UBPR CUBPR UPL RecVAE DCF ORACLE CDR SDR UIDR
Builds on: uPU nnPU RMF UBPR UPL
Closest papers by meaning
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.85)
- Plug-In Diffusion Model for Sequential Recommendation (AAAI 2024 · similarity 0.85)
- CURE4Rec: A Benchmark for Recommendation Unlearning with Deeper Influence (NeurIPS 2024 · similarity 0.85)
- Adaptive Hardness Negative Sampling for Collaborative Filtering (AAAI 2024 · similarity 0.85)
- ImplicitSLIM and How it Improves Embedding-based Collaborative Filtering (ICLR 2024 · similarity 0.85)
- Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (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.