Causal Debiasing in Ant Group Marketing Recommendation: Data Fusion and Backdoor Adjustment

Introduction: This article presents causal debiasing methods applied to Ant Group’s marketing recommendation scenarios, covering background, common biases, causal graph analysis, and practical solutions.

Background: Recommendation loops can introduce selection, exposure, and popularity biases because models are trained on observational data affected by exposure strategies and user choices, leading to a gap between empirical risk and true risk.

Causal Debiasing: Two main approaches are described. (1) Data‑fusion debiasing using a meta‑learning based MDI model that leverages both unbiased (teacher) data and biased data with augmentation, trained in two stages and optimized with IPS and IMP losses. (2) Backdoor‑adjustment debiasing that removes spurious paths in the causal graph by applying do‑calculus, effectively cutting the D→U and K→I paths.

Data‑Fusion Model: The MDI framework pre‑trains an unbiased teacher model, copies its parameters to a fused model, updates with biased plus augmented data, and iteratively refines until convergence. Experiments on Yahoo R3, Coat, and an internal Ant dataset show that MDI outperforms baselines such as IPW, AutoDebias, PFDR, and DoubleRobust in MAE and MSE.

Backdoor Adjustment: By modeling the causal graph with user (U), item (I), interaction (Y), and policy variables (T), the method blocks backdoor paths (U‑D‑T‑Y, I‑K‑T‑Y) and estimates the true causal effect using do‑operators. Evaluations on public coupon datasets and real‑world marketing data demonstrate superior AUC compared with IPW, Unawareness, FairCo, MACR, PDA, and DecRS.

Application in Ant Group: These causal debiasing techniques are deployed in advertising and e‑commerce marketing scenarios to mitigate rule‑based exposure constraints, protect small merchants, and improve user experience, resulting in measurable gains in redemption rates and sales.

Conclusion: Causal debiasing via data fusion and backdoor adjustment provides robust performance across varying amounts of unbiased data, and has been successfully applied in production at Ant Group.