RankUplift: Ranking-Aware Two-Stage Learning for Optimal Targeted Interventions

Uplift modeling aims to identify individuals who respond most positively to an intervention. While numerous methods have been proposed to estimate individual treatment effects (ITEs), most optimize for pointwise prediction accuracy, which misaligns with the practical goal of ranking individuals by responsiveness to enable effective targeting under budget constraints. Moreover, real-world observational data often suffer from selection bias and poor overlap, further degrading ranking quality when models prioritize absolute error over ordinal fidelity. To bridge this gap, we propose RankUplift, a two-stage framework that decouples causal effect estimation from ranking optimization. In the first stage, a robust ITE estimator generates surrogate labels; in the second stage, a scoring function is trained via the listwise ListNet objective to preserve global uplift ranking structure. We prove that under mild conditions, RankUplift asymptotically recovers the true ITE ranking and yields near-optimal top-k policies. Extensive experiments across benchmark and semi-synthetic datasets demonstrate consistent improvements over state-of-the-art baselines in both ranking metrics and practical targeting performance, confirming that aligning the learning objective with decision-making needs leads to superior real-world outcomes.