Research map / Probabilistic and scientific modeling / Molecules and proteins
Backdoor Adjustment via Group Adaptation for Debiased Coupon Recommendations
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 BAGA, a backdoor-adjustment training paradigm that segments users and coupons into groups and models them with a dual-tower CGC multi-task framework to remove matching bias in coupon recommendation CTR models.
Problem
Marketing tactics that interfere with coupon allocation create imbalanced user-coupon interaction data (matching bias), which confounds conversion prediction and yields spurious correlations in coupon recommendation.
Approach
Treat user/coupon attributes tied to marketing tactics as confounders, apply backdoor adjustment by discretizing them into adaptive user and coupon groups, and train group-wise sub-models efficiently via a dual-tower CGC multi-task architecture with O(J+K) complexity instead of O(J*K).
Key results
- On the simulated unbiased test sets, BAGA on DeepFM reaches AUC 0.710/0.749/0.681/0.869/0.933 across Ds A-E, highest among compared methods on all five datasets.
- Friedman test (tauF 50.9999 > critical 2.1620) and Bonferroni-Dunn CD diagram show BAGA has the lowest average rank in AUC.
- Online A/B test on a real coupon campaign: use rate 4.83% vs 4.76% baseline (+1.54%) and coupons used 70,673 vs 69,556 (+1.60%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply BAGA debiasing | DEEPFM · Ds A | AUC: 0.697±0.002 -> 0.710±0.002 | Helped |
| apply BAGA debiasing | DEEPFM · Ds B | AUC: 0.726±0.004 -> 0.749±0.004 | Helped |
| apply BAGA debiasing | DEEPFM · Ds C | AUC: 0.653±0.003 -> 0.681±0.002 | Helped |
| apply BAGA debiasing | DEEPFM · Ds D | AUC: 0.839±0.003 -> 0.869±0.004 | Helped |
| apply BAGA debiasing | DEEPFM · Ds E | AUC: 0.919±0.002 -> 0.933±0.003 | Helped |
| use the CGC architecture instead of multi-task MLP layers (ablation) | BAGA with DEEPFM · Ds A | AUC: 0.697±0.001 -> 0.710±0.002 | Helped |
| use the CGC architecture instead of multi-task MLP layers (ablation) | BAGA with DEEPFM · Ds B | AUC: 0.724±0.002 -> 0.749±0.004 | Helped |
| use the CGC architecture instead of multi-task MLP layers (ablation) | BAGA with DEEPFM · Ds C | AUC: 0.663±0.005 -> 0.681±0.002 | Helped |
| use the CGC architecture instead of multi-task MLP layers (ablation) | BAGA with DEEPFM · Ds D | AUC: 0.836±0.004 -> 0.869±0.004 | Helped |
| use the CGC architecture instead of multi-task MLP layers (ablation) | BAGA with DEEPFM · Ds E | AUC: 0.921±0.002 -> 0.933±0.003 | Helped |
| apply BAGA debiasing | DEEPFM · Online marketing coupon recommendation | Use rate: +1.54% | Helped |
| apply BAGA debiasing | DEEPFM · Online marketing coupon recommendation | Coupons used: +1.60% | Helped |
Surprising findings
- Simple re-weighting (IPW) yields performance comparable to the vanilla model, suggesting it cannot precisely correct matching bias.
- BAGA has longer training and testing time than baselines, though the authors describe it as acceptable.
Limitations
- Modeling all user-coupon group pairs is O(J*K); the dual-tower approximation is needed for scalability.
- BAGA does not significantly outperform some baselines in pairwise t-tests on several datasets.
- Four of five evaluation datasets are filtered to simulate marketing tactics, and the main dataset is proprietary and not released.
Experimental setup
Datasets: Amazon dataset Ele.me data (Tianchi 2022) MovieLens Alibaba Ads Click dataset (Tianchi 2018) proprietary marketing coupon dataset (Ds E)
Benchmarks: Amazon Ele.me MovieLens Alibaba Ads Click proprietary coupon dataset Ds E
Metrics: AUC use rate number of coupons used training/testing time
Baselines: IPW FAIRCO PDA DICE DMBR BAGA-mlp
Builds on: backdoor adjustment (Pearl 2009) Customized Gate Control (CGC) DMBR
Closest papers by meaning
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.84)
- Invariant Deep Uplift Modeling for Incentive Assignment in Online Marketing via Probability of Necessity and Sufficiency (ICML 2025 · similarity 0.84)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.83)
- TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems (NeurIPS 2025 · similarity 0.83)
- Debiased Collaborative Filtering with Kernel-Based Causal Balancing (ICLR 2024 · similarity 0.83)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.82)
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.