Research map / Label-efficient and robust learning / Recognition and long-tail
Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering
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 User-DR, Item-DR and User-Item-DR doubly robust estimators for debiased collaborative filtering that stay unbiased under user- and/or item-specific pseudo-labeling biases, plus an attention-weighted propensity reconstruction learning procedure.
Problem
Doubly robust debiasing in recommender systems is only unbiased when pseudo-labelings (imputed ratings) are exactly accurate or propensities are accurate, which is hard to satisfy in practice under sampling selection bias.
Approach
Constrain the learned propensity model so that the weighted imputation error sums to zero per user and/or per item; this makes the DR estimator unbiased when pseudo-labels carry arbitrary user-specific and/or item-specific inductive bias, and an attention mechanism adaptively sets constraint weights while reconstructing inverse propensities closest to the base ones to control variance.
Key results
- On semi-synthetic ML-100K, UIDR achieves the lowest relative error in most bias settings (e.g., 0.003 vs DR's 0.047 in the ONE setting with user-specific bias).
- UDR reaches AUC 0.739 on COAT and 0.802 on KuaiRec versus best baseline TDR at 0.730 and 0.792, respectively.
- Ablation on KuaiRec shows removing the constraint loss (UDR w/o Lre) drops AUC from 0.802 to 0.787 and removing the propensity-reconstruction weight gamma gives 0.754 AUC, approximately DR-JL level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Coat unbiased test set | AUC: 0.730±0.008 -> 0.740∗±0.006 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Coat unbiased test set | NDCG@5: 0.656±0.011 -> 0.722∗±0.006 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Coat unbiased test set | F1@5: 0.493±0.010 -> 0.539∗±0.005 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Music unbiased test set | AUC: 0.702±0.002 -> 0.713∗±0.001 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Music unbiased test set | NDCG@5: 0.680±0.004 -> 0.752∗±0.001 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · Music unbiased test set | F1@5: 0.337±0.002 -> 0.382∗±0.001 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · KuaiRec unbiased test set | AUC: 0.793±0.004 -> 0.804∗±0.004 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · KuaiRec unbiased test set | NDCG@50: 0.799±0.003 -> 0.804∗±0.004 | Helped |
| use UIDR instead of the strongest baseline | MF-based debiased collaborative filtering · KuaiRec unbiased test set | F1@50: 0.606±0.001 -> 0.610∗±0.003 | Helped |
| remove γ from the propensity reconstruction loss (ablation) | UDR · KuaiRec unbiased test set | AUC: 0.802 -> 0.754 | Hurt |
| remove γ from the propensity reconstruction loss (ablation) | UDR · KuaiRec unbiased test set | NDCG@50: 0.804 -> 0.754 | Hurt |
| remove γ from the propensity reconstruction loss (ablation) | UDR · KuaiRec unbiased test set | F1@50: 0.610 -> 0.589 | Hurt |
Surprising findings
- Item-DR performs slightly worse than User-DR and User-Item-DR on real datasets, attributed to item-side inductive bias being smaller than user-side in practice.
- Simple constant or MLP-based constraint weights underperform the attention mechanism for lambda_u, with MLP prone to overweighting specific constraints.
Limitations
- The learning method cannot guarantee all unbiasedness constraints hold strictly due to computational burden.
- Unbiasedness relies on the inductive bias being user-specific and/or item-specific; general biases are only partially reduced.
- Semi-synthetic experiments rely on simulated propensities and biased predictions.
Experimental setup
Datasets: MovieLens 100K COAT MUSIC KuaiRec
Benchmarks: ML-100K semi-synthetic COAT MUSIC KuaiRec
Metrics: relative error AUC NDCG@K F1@K
Baselines: MF IPS ASIPS DR DR-JL MRDR-JL CVIB DIB DR-MSE MR TDR TDR-JL
Builds on: doubly robust estimator DR-JL MRDR error-imputation-based learning inverse propensity scoring
Closest papers by meaning
- Debiased Collaborative Filtering with Kernel-Based Causal Balancing (ICLR 2024 · similarity 0.87)
- Adaptive Hardness Negative Sampling for Collaborative Filtering (AAAI 2024 · similarity 0.87)
- From Pairwise to Ranking: Climbing the Ladder to Ideal Collaborative Filtering with Pseudo-Ranking (AAAI 2025 · similarity 0.86)
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.86)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.85)
- Fine-Grained Dynamic Framework for Bias-Variance Joint Optimization on Data Missing Not at Random (NeurIPS 2024 · 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.