Research map / Probabilistic and scientific modeling / Treatment effects
Alleviating Dual Biases in Recommendation (Student Abstract)
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 DCR, a disentangled counterfactual reasoning framework that separates biased (popularity and conformity) and unbiased embeddings for LightGCN-based recommenders and debiases both indirect and direct effects at inference.
Problem
Existing causal-inference debiasing methods for recommendation address either popularity bias or user conformity bias, and conflate biased and unbiased representations, so dual biases are not comprehensively mitigated.
Approach
Disentangle user and item embeddings into biased and unbiased parts using mutual information minimization/maximization, train with a causal graph including item popularity and user conformity, and at inference suppress biased embeddings (indirect effects) and subtract a counterfactual popularity-conformity score (direct effects).
Key results
- DCR achieves Recall 0.2418 and NDCG 0.0924 on Adressa, outperforming the best baseline PPAC (0.2157 / 0.0797) by 12.1% and 15.9%.
- On Gowalla DCR reaches Recall 0.1950 vs. 0.1849 for PPAC (5.5% improvement) on unbiased test sets.
- DCR improves Recall on Yelp2018 by 6.8% over the best existing method (0.1073 vs 0.1005).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCR instead of PPAC | LightGCN recommendation on Adressa · Adressa | Recall: 0.2157 -> 0.2418 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on Adressa · Adressa | NDCG: 0.0797 -> 0.0924 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on MovieLens-1M · MovieLens-1M | Recall: 0.4023 -> 0.4395 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on MovieLens-1M · MovieLens-1M | NDCG: 0.2473 -> 0.2756 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on Gowalla · Gowalla | Recall: 0.1849 -> 0.1950 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on Gowalla · Gowalla | NDCG: 0.0771 -> 0.0810 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on Yelp2018 · Yelp2018 | Recall: 0.1005 -> 0.1073 | Helped |
| Use DCR instead of PPAC | LightGCN recommendation on Yelp2018 · Yelp2018 | NDCG: 0.0401 -> 0.0429 | Helped |
Surprising findings
- MACR, despite targeting dual biases, underperforms single-bias methods such as PPAC, which the authors attribute to failing to separate true interests from bias-influenced behavior.
Limitations
- Evaluated only with LightGCN as the base recommender
- Relies on constructed unbiased test sets rather than real unbiased data
- Hyperparameter lambda must be tuned
Experimental setup
Datasets: Adressa MovieLens-1M Gowalla Yelp2018
Benchmarks: Adressa MovieLens-1M Gowalla Yelp2018
Metrics: Recall NDCG
Baselines: UDIPS DICE MACR PPAC
Builds on: LightGCN MACR DICE PPAC
Closest papers by meaning
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.87)
- CoDeR: Counterfactual Demand Reasoning for Sequential Recommendation (AAAI 2025 · similarity 0.87)
- Debiased Collaborative Filtering with Kernel-Based Causal Balancing (ICLR 2024 · similarity 0.86)
- Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering (ICML 2024 · similarity 0.86)
- TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems (NeurIPS 2025 · similarity 0.85)
- Causal Representation Learning for Generalisable Recommendation (NeurIPS 2026 · 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.