Mechanism-aware personalized learning intervention for short- and long-term academic outcome modeling

Improving academic achievement while also modeling students' longer-horizon academic development has become an important goal in educational psychology and AI-supported learning research. However, adaptive learning, learner modeling, feedback generation, and academic prediction are often studied separately, and the learner conditions used to guide intervention are not always defined with sufficient theoretical and methodological transparency.

This study proposes MAPLe-I (Mechanism-Aware Personalized Learning Intervention), a unified framework that integrates learner-state encoding, theory-informed latent proxy estimation, personalized intervention policy learning, and dual-horizon academic outcome modeling. Learner state is encoded from graph-structured knowledge relations and temporal behavioral sequences, and the model estimates three intervention-relevant latent proxies: engagement-related state, self-regulation-related readiness, and affective support need. Because the public datasets do not include observed intervention decisions, policy supervision was implemented using rule-derived pedagogical pseudo-labels. The framework was evaluated offline on ASSISTments 2012-2013 with Affect, EdNet-KT1, and OULAD against non-personalized digital intervention, traditional adaptive learning, AI-based feedback, and learner-modeling or academic-prediction baselines.

MAPLe-I achieved stronger comparative benchmark performance across multiple settings. On ASSISTments, it obtained an AUC of 0.824 and an Intervention Alignment Rate of 0.785. On EdNet-KT1, it obtained an AUC of 0.812. On OULAD, it achieved a Macro-F1 of 0.772 and reduced RMSE to 0.315. Ablation and sensitivity analyses further supported the contribution of the temporal encoder, mechanism head, intervention policy head, and dual-horizon optimization strategy.

The findings suggest that integrating learner-state diagnosis, theory-informed latent proxies, and intervention-policy modeling can improve offline predictive and rule-alignment performance in benchmark settings. However, the mechanism variables are computational proxies, the intervention labels are rule-derived rather than observed, and the results do not establish causal effects or authentic classroom impact. MAPLe-I should therefore be interpreted as a transparent benchmark framework for mechanism-aware personalized learning rather than as a validated real-world intervention system.