Backdoor Adjustment via Group Adaptation for Debiased Coupon Recommendations

Junpeng Fang, Gongduo Zhang, Qing Cui, Caizhi Tang, Lihong Gu, Longfei Li, Jinjie Gu, Jun ZhouAAAI 2024 · DOI

Introduction: The Challenge of Matching Bias

In modern e-commerce, coupon recommendation systems are vital for driving user engagement. However, these systems often inherit "matching bias" from historical data. This bias arises because marketing tactics—such as offering specific coupons only to certain user segments—create artificial correlations. Standard models trained on this data learn to replicate these tactics rather than identifying genuine user preferences, leading to suboptimal recommendations and inefficient marketing spend.

Figure 1: Illustration of matching bias where marketing tactics (e.g., public transport coupons for car-less users) distort predicted conversion probabilities.

Methodology: A Causal Inference Approach

The proposed framework, BAGA (Backdoor Adjustment via Group Adaptation), addresses matching bias through the lens of causal inference. The researchers identify marketing tactics as a confounding variable $T$ (representing hierarchical attributes of users and coupons). This confounder influences both the treatment (the matching of user $U$ and item $I$) and the outcome (conversion $Y$). This relationship forms a "backdoor path" ($U, I \leftarrow T \rightarrow Y$) that introduces spurious correlations into the training data.

Figure 3: Causal diagrams representing (a) the confounding effect of matching bias and (b) the backdoor adjustment strategy used to block the influence of confounder T.

The BAGA Framework

To eliminate this bias efficiently, BAGA implements several key components:

  1. Backdoor Adjustment: Utilizing the "do-operator" from causal calculus, the model mathematically intervenes to block the influence of $T$, allowing for the estimation of the true causal effect of a coupon on a user.
  2. Adaptive Group Segmentation: Users and coupons are partitioned into $J$ and $K$ groups based on the attributes used in marketing tactics.
  3. Dual-Tower Multi-Task Architecture: To avoid the $O(J \times K)$ complexity of modeling every possible group pair, BAGA uses a dual-tower approach. This reduces complexity to $O(J + K)$, making it suitable for large-scale production.
  4. Customized Gate Control (CGC): The model employs a CGC architecture with both task-specific and shared experts. This design enables the model to capture group-specific nuances while benefiting from global data patterns, which is particularly effective for minority groups with limited data.

Experimental Evaluation

BAGA was tested on five diverse datasets: Amazon (Ds_A), Ele.me (Ds_B), MovieLens (Ds_C), Alibaba Ads (Ds_D), and a large-scale proprietary dataset from Ant Group (Ds_E) containing over 60 million samples.

Key Results

  • State-of-the-Art Performance: BAGA outperformed several competitive debiasing methods, including IPW, DICE, and DMBR, across all datasets in terms of Area Under Curve (AUC).
  • Statistical Significance: Rigorous testing (Friedman and Bonferroni-Dunn) confirmed that BAGA's improvements are statistically significant.
  • Online A/B Testing: In a real-world deployment on a major payment platform, BAGA achieved a 1.54% increase in coupon use rate and a 1.60% increase in total coupons used compared to a strong DeepFM baseline.

Conclusion

By framing marketing tactics as causal confounders, BAGA provides a principled way to de-bias coupon recommendations. Its efficient dual-tower architecture and shared-expert learning mechanism make it a practical and powerful solution for large-scale e-commerce platforms seeking to optimize marketing efficiency and user satisfaction.

Summary updated

BibTeX
@inproceedings{fang2024backdoor,
  author = {Junpeng Fang and Gongduo Zhang and Qing Cui and Caizhi Tang and Lihong Gu and Longfei Li and Jinjie Gu and Jun Zhou},
  title = {{Backdoor Adjustment via Group Adaptation for Debiased Coupon Recommendations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i11.29081},
}
Figure