Causal Representation Learning for Personalized Adaptive Learning Path Optimization

Personalized adaptive learning systems commonly optimize learning paths using correlational representations derived from observational logs, which can internalize confounding and degrade under cohort or course shifts. This study introduces a unified framework for causal representation learning that jointly learns learner-state embeddings and optimizes sequencing policies under offline constraints. Experiments used LMS traces from 12 course offerings comprising 12,438 learners and 1.86 million interaction events, yielding 1,734,920 cleaned decision points. Descriptive analysis revealed a mid-semester attrition concentration, with the sharpest contraction in active learners between Weeks 4 and 7 and a semester-level dropout proxy rate of 17.6 percent. Causal effect estimation showed substantial treatment heterogeneity for difficulty assignment: the estimated effect of Hard versus Medium activities on next-step mastery ranged from minus 0.028 in the lowest baseline-skill quartile to plus 0.031 in the highest quartile, while the aggregate average treatment effect was near zero at plus 0.004 due to cancellation. The proposed representation improved next-step mastery predictability to an AUC of 0.83 while reducing environment leakage, with course-domain predictability declining to 0.46. Offline policy evaluation demonstrated consistent uplift across estimators, increasing doubly robust value from 0.407 under the logged baseline to 0.463 under the causal-representation policy, while maintaining higher support alignment with an effective sample size of 0.81 versus 0.73 for a correlational policy. Deployment-facing indicators improved concurrently, including retention proxy (10.8 active weeks versus 9.6) and reduced pacing instability (difficulty jump rate 8.4 percent versus 10.9 to 12.1 percent under ablations). Ablation studies confirmed that invariance constraints, counterfactual rollouts, and stable-dynamic disentanglement are structurally necessary to preserve both mastery and persistence benefits.