From Prediction to Prescription: A Causal-AI for Personalized Learning Path Optimization in MOOCs

Abstract

Massive Open Online Courses (MOOCs) generate vast amounts of learner interaction data, enabling predictive models to identify at-risk students. However, most existing approaches stop at prediction, offering limited guidance on which interventions should be applied to improve outcomes. This paper proposes a Causal-AI–driven reinforcement learning (RL) framework that transitions from prediction to prescription by integrating causal effect estimation with policy optimization. Specifically, uplift modeling is employed to estimate heterogeneous treatment effects of candidate learning actions, and these causal signals are embedded into an Actor–Critic RL agent through causal regularization. The resulting framework prescribes adaptive intervention policies that are both data-driven and causally grounded. We evaluate the approach on the OULAD dataset. Baseline predictive models achieve strong accuracy (AUC = 0.795) and calibration, while uplift distributions reveal heavy-tailed heterogeneity across learners. Off-policy evaluation demonstrates that the causal-RL policy consistently outperforms the observed behavior and standard RL baselines (Greedy Uplift SNIPS ≈ 0.619 vs. baseline 0.589). Subgroup analysis highlights disparities across age and gender, motivating fairness-aware extensions. A case study further illustrates how adaptive recommendations can redirect learner trajectories toward success. These findings demonstrate that causal-RL integration enables prescriptive, interpretable, and equitable personalization in MOOCs.