Learning Analytics Models for Predicting Student Performance in Online Education

Abstract

Learning analytics leverages the vast behavioural and performance data generated by online learning platforms to predict student outcomes, identify at-risk learners, and enable timely instructional interventions. Accurate early prediction of student performance -- dropout risk, final grade, and course completion -- is a high-value application with direct impact on learner success rates in online education, where dropout rates frequently exceed 50% in MOOCs and 20-30% in online degree programmes. This paper proposes the Learning Analytics Predictive Model Framework (LAPMF), a systematic evaluation of six predictive modelling approaches -- logistic regression, random forest, gradient boosting (XGBoost), LSTM sequence models, transformer-based sequence models, and graph neural networks -- across four prediction tasks on three large-scale online education datasets (Coursera, edX Open, and Open University Learning Analytics Dataset, combined N = 84,200 students). LAPMF introduces the Predictive Analytics Utility Score (PAUS) integrating prediction accuracy, early warning timeliness, and model interpretability for educational deployment. Key results: XGBoost achieves the highest overall PAUS (0.902) with AUC = 0.924 for dropout prediction; LSTM achieves the highest early-warning accuracy at week 2 (AUC = 0.884); GNN achieves the best performance prediction for collaborative learning contexts (AUC = 0.896). The framework provides model selection guidance and an open benchmark suite for educational predictive analytics.