📝 SummarXiv

Abstract

Uplift modeling is a key technique for promotion optimization in recommender systems, but standard methods typically fail to account for interference, where treating one item affects the outcomes of others. This violation of the Stable Unit Treatment Value Assumption (SUTVA) leads to suboptimal policies in real-world marketplaces. Recent developments in interference-aware estimators such as Additive Inverse Propensity Weighting (AddIPW) have not found their way into the uplift modeling literature yet, and optimising policies using these estimators is not well-established. This paper proposes a practical methodology to bridge this gap. We use the AddIPW estimator as a differentiable learning objective suitable for gradient-based optimization. We demonstrate how this framework can be integrated with proven response transformation techniques to directly optimize for economic outcomes like incremental profit. Through simulations, we show that our approach significantly outperforms interference-naive methods, especially as interference effects grow. Furthermore, we find that adapting profit-centric uplift strategies within our framework can yield superior performance in identifying the highest-impact interventions, offering a practical path toward more profitable incentive personalization.