Personalized Knowledge Tracing for Student Performance Prediction in Digital Learning Environments

Abstract

Digital learning platforms generate substantial amounts of behavioral data that can be used to understand student learning progress and identify potential academic difficulties. However, conventional performance prediction methods frequently overlook individual learning trajectories and differences in knowledge acquisition. This paper introduces a personalized knowledge tracing framework that integrates sequential learning interactions, concept dependencies, and learner-specific representations for student performance prediction. The proposed architecture combines attention-based sequence modeling with a hierarchical knowledge representation network to capture both short-term learning behaviors and long-term concept mastery. The illustrative evaluation considers 8.6 million learning interactions from approximately 124,000 students enrolled in 38 online courses. The proposed framework achieves an AUROC of 0.862 for next-response prediction, compared with 0.817 for a conventional knowledge tracing baseline. Prediction calibration improves substantially, with expected calibration error decreasing from 0.068 to 0.031. At a fixed false-positive rate of 15%, the model identifies 76% of students classified as academically at risk, compared with 64% for the baseline approach. These results suggest that personalized sequence representations can improve educational performance modeling and support more targeted learning interventions.