LEAP: Dynamic Graph-Enhanced Learning Path Quality Assessment for Personalized Adaptive Education

Adaptive educational systems aim to deliver personalized learning [1, 2], but dynamically assessing the quality of a student's learning path-beyond pointwise anomaly detection-remains a key challenge. Inspired by recent advances in adaptive learning schemes [3] and federated neural architectures [4], our approach builds upon these foundational works while extending them to the educational domain. To tackle this, we propose LEAP (LEArning Path quality assessment and Personalization), a novel framework for real-time, personalized learning path quality assessment. LEAP incorporates a Dynamic Collaborative Graph Encoder (DCGE) to model evolving representations of learners and knowledge points. Central to the framework is the Path Quality Synthesizer (PQS), which breaks down path quality into three interpretable, educationally-grounded dimensions: topological coherence in a knowledge graph, difficulty adaptation to learner proficiency, and mastery prediction error. For scalable streaming, a Robust Adaptive Isolation Forest (RAIF) module serves as an efficient pre-filter. Evaluations on three real-world educational datasets (ASSISTments 2017, EdNet KT1, Junyi Academy) show that LEAP outperforms strong baselines by up to 7.2% in ROC-AUC and 13.3% in Average Precision. Ablation studies verify each component's importance: the DCGE enables dynamic representations, the PQS is vital for accurate assessment, and the RAIF reduces computational cost by 8.7× with minimal accuracy loss. LEAP thus offers a robust, interpretable, and efficient solution for guiding personalized interventions in adaptive education.