Primary Submission Category: Machine Learning and Causal Inference
Screens to Smarts: Regularized Apprenticeship Learning with Attention for Inferring Behavioral Pathways Linked to Health Knowledge Gains in a Digital Intervention
Authors: Rahul Ladhania, Rema Padman, Shreyas Vajjhala,
Presenting Author: Rahul Ladhania*
Gamified digital interventions are increasingly used to promote health knowledge and behavior, yet limited evidence exists on which in-intervention behavioral pathways are linked to downstream learning outcomes. We analyze gameplay telemetry from an 11-week school-based digital health intervention in India to discover interpretable gameplay strategies associated with improvements in children’s health knowledge. We model gameplay trajectories as behavioral sequences, and develop a novel inverse reinforcement learning framework that incorporates L1 regularization and an attention mechanism to extract interpretable latent behavioral patterns from high-dimensional, small-sample telemetry data. Survey-based knowledge gains serve as a downstream reward signal to recover latent gameplay strategies associated with improved learning outcomes. We find that students who more consistently engage with health-aligned mechanics (e.g., exercising, using power-ups effectively) exhibit larger knowledge gains. Incorporating regularization and attention substantially improves alignment between inferred behavioral patterns and observed outcomes, yielding transparent representations that highlight the most behaviorally relevant gameplay features. Our findings demonstrate how telemetry-driven behavioral modeling can support mechanism discovery and hypothesis generation, informing design of future experiments, adaptive interventions, and personalization strategies in digital health and education.