Causal Debiasing Techniques for Recommendation and Marketing Scenarios

The article introduces causal debiasing methods applied to recommendation and marketing systems, beginning with an overview of bias sources such as selection bias, exposure bias, and popularity bias that arise from training on observational data rather than ideal data.

It explains the concept of causal debiasing, illustrating how confounders like smoking in a health example create spurious correlations, and describes two main strategies: using unbiased data for representation learning and adjusting observational data via causal graphs.

Section 2 details a data‑fusion based debiasing approach called the MDI model, which leverages both unbiased and biased data through meta‑learning. The training pipeline consists of pre‑training an unbiased teacher model f_u , copying its parameters to a biased model f_d , updating an augmentation model f_p on augmented data, and iteratively refining f_d until convergence.

Experimental results on public datasets (Yahoo R3, Coat) and an internal e‑commerce dataset show that MDI consistently outperforms baselines such as IPS, Double Robust, AutoDebias, and PFDR in terms of MAE, MSE, and AUC, especially when the proportion of unbiased data varies.

Section 3 introduces a back‑door adjustment method that removes spurious paths in the causal graph (e.g., U‑D‑T‑Y and I‑K‑T‑Y) by cutting the D→U and K→I connections, allowing the model to estimate the true causal effect of user and item representations on the outcome.

Experiments on synthetic and real marketing datasets (Tianchi, 84.51 coupons) demonstrate that the back‑door adjusted model (DMBR) achieves higher AUC and conversion rates compared with methods like IPW, Unawareness, FairCo, MACR, PDA, and DecRS.

Finally, the article discusses practical deployments of these causal debiasing techniques within Ant Group’s advertising and e‑commerce marketing platforms, where rule‑based constraints and strategy biases are common, and reports online improvements in coupon redemption rates and sales volume.