Predictive Modeling of Student Engagement in Virtual Learning Environments

Abstract

Student engagement in virtual learning environments (VLEs) -- the cognitive, behavioural, and emotional investment learners bring to online educational activities -- is the strongest modifiable predictor of academic success, with disengagement typically preceding dropout by 2-4 weeks. Accurate predictive models of VLE engagement enable timely proactive interventions that can re-engage at-risk learners before disengagement becomes withdrawal. This paper proposes the Student Engagement Predictive Model Framework (SEPMF), a systematic evaluation of six predictive modelling approaches -- multivariate time-series regression, gradient boosting (XGBoost), bidirectional LSTM, attention-based transformer, multimodal fusion (behavioural + affective + cognitive signals), and graph attention networks -- for predicting three engagement dimensions (behavioural, cognitive, and emotional) across 58,400 learner sessions from four VLE platforms. SEPMF introduces the Engagement Prediction Utility Score (EPUS) integrating prediction accuracy, early detection lead time, and intervention actionability. Key results: multimodal fusion achieves the highest EPUS (0.918) with engagement state AUC = 0.936; bidirectional LSTM achieves the earliest detection (2.4 weeks before disengagement); behavioural features alone achieve AUC = 0.892 without affective sensing hardware. The framework provides engagement prediction model selection guidance and an open benchmark suite.