Cross-Head Attention Uplift Network with Inverse Propensity Score under Unobserved Confounding

Uplift modeling aims to estimate the Individual Treatment Effect (ITE)—the difference in outcomes for a single unit under treatment versus control. A primary challenge in this field is that researchers only observe one outcome per unit (the factual), leaving the other (the counterfactual) unknown. Furthermore, observational data often suffer from selection bias due to unobserved confounders, which distort the relationship between treatment assignment and outcomes. This paper seeks to improve ITE prediction by better capturing inter-group similarities and robustly debiasing models when unobserved confounders are present.

The authors propose two main innovations. First, the Cross-Head Attention Uplift Network (CHAUN) uses a dual-branch architecture to process treatment and control groups. Instead of treating these branches as independent, CHAUN employs a cross-head attention mechanism that allows the model to dynamically exchange information between the two branches, effectively leveraging inter-group correlations to refine representations. Second, to handle unobserved confounding, the authors introduce Robust Adversarial Inverse Propensity Score (RA-IPS). While standard propensity score weighting assumes all confounders are observed, RA-IPS treats the true propensity score as existing within an uncertainty set and adversarially optimizes the weights to minimize the impact of worst-case unobserved confounding.

CHAUN demonstrates superior performance compared to existing state-of-the-art uplift models across public datasets (CRITEO-UPLIFT, LAZADA) and a proprietary e-commerce dataset, achieving relative improvements of up to 25.6% in QINI scores. The RA-IPS method further improves model robustness, outperforming standard inverse propensity score (IPS) weighting by 5.4% in scenarios with unobserved confounding. The authors also provide a theoretical proof that, given access to true propensity scores, ITE remains identifiable even when unobserved confounders are present.