Learning framework with joint task optimization applied to consumer health applications with behavioral nudges
Shaudi Mahdavi Hosseini · 2020
We consider a consumer-facing healthcare mobile app that recommends optimal user behaviors based on the user's current and historical engagement with the application, including both inferred and explicit preferences. The app separately pushes intermittent behavioral nudges to elicit specific optimal user health behaviors like messaging a healthcare provider or recording one's meals after extended disengagement. We propose applying a reinforcement learning approach, and jointly optimizing the recommendation of a menu of items and sending the user behavioral nudges in a multi-stage paradigm. We propose this as an alternative to optimizing the recommendations and nudges separately or in sequence, as is the prevailing approach across different but analogous e-commerce recommender system settings. We argue that this solution approach optimizes user welfare in the short- and long-term across generalized reward functions.