Optimizing User Engagement through Adaptive Ad Sequencing

In this paper, we propose a unified dynamic framework for adaptive ad sequencing that optimizes user engagement with ads. Our framework comprises three components -- (1) a Markov Decision Process that incorporates inter-temporal trade-offs in ad interventions, (2) an empirical framework that combines machine learning methods with insights from causal inference to achieve personalization, counterfactual validity, and scalability, and (3) a robust policy evaluation method. We apply our framework to large-scale data from the leading in-app ad network of an Asian country. We find that the dynamic policy generated by our framework improves the current practice in the industry by 5.76%. This improvement almost entirely comes from the increased average ad response to each impression instead of the increased usage by each user. We further document a U-shaped pattern in improvements across the length of the user's past history, with high values when the user is new or when enough data are available for the user. Next, we show that ad diversity is higher under our policy because it manages users' attention more effectively. We conclude by discussing the implications and broad applicability of our framework to settings where a platform wants to sequence content to optimize user engagement