Building a User-Level Targeting Platform

Many consumer brands utilize coupons and other monetary incentives to grow their business. Most utilize a one-size-fits-all approach to coupons that does not account for differences in their customers or changes over time. While (contextual) reinforcement learning provides an approach for learning the optimal coupon to give each user, it treats each action (coupon) as independent. Heterogeneous treatment effect (HTE) models can help account for the relationship between actions and provide an efficient way of estimating user-level responses to coupons. We will explain our method of combining reinforcement learning and HTE modeling to develop a user-level targeting platform that can help choose the right coupon for each user. This platform can take in inputs on objective functions and constraints and balance explore/exploit trade-offs to learn the optimal treatment. We modify existing state-of-the-art HTE models so that we can pool data across multiple coupons and efficiently explore brand new options. This talk will walk through the structure of the platform, dive into the statistical challenges and discuss our solutions.