Causal Debiasing Methods for Ant Group's Marketing Recommendation Scenarios

The article introduces causal debiasing techniques applied to Ant Group's marketing recommendation scenarios, beginning with an overview of bias generation in closed‑loop recommendation systems and describing three typical biases: selection, exposure, and popularity.

It then explains causal debiasing, illustrating how confounders such as smoking (a common cause) can create spurious associations, and how do‑calculus can be used to remove these false links.

Two main correction methods are detailed:

Data‑fusion based debiasing (MDI model) : a meta‑learning framework that jointly leverages unbiased teacher data and biased data, training in two stages—pre‑training an unbiased teacher model on unbiased data, then iteratively updating a fusion model with weighted samples from both sources. Experiments on Yahoo R3, Coat, and an internal dataset show superior MSE/MAE performance compared to baselines such as IPS, Double‑Robust, and PFDR.

Back‑door adjustment debiasing (DMBR model) : constructs a causal graph for recommendation, identifies back‑door paths (e.g., U‑D‑T‑Y, I‑K‑T‑Y), and cuts them by modeling an unbiased representation T, enabling true causal inference. Evaluations on Tianchi, 84.51, and a real e‑commerce marketing dataset demonstrate higher AUC and better conversion rates than methods like IPW, Unawareness, FairCo, MACR, PDA, and DecRS.

The paper also discusses ablation studies confirming the necessity of the augmentation module and sensitivity analyses showing robustness across varying proportions of unbiased data.

Finally, practical applications within Ant Group are highlighted, where causal debiasing is employed in advertising and e‑commerce marketing to mitigate rule‑based or strategy‑driven biases, improving coupon redemption rates and overall user experience.

The article concludes with a thank‑you note.