Early Prediction of Student Dropout Risk from Learning Management System Interaction Sequences using Recurrent Deep Learning Models

Given the fluidity of engagement in online and blended learning, early identification of students who are at risk of dropping out remains a major challenge. The study aims to explore the effectiveness of recurrent deep learning models in predicting students' dropout using sequential interaction data from the LMS. An interaction dataset from a public LMS system with approximately 100,000 activity logs from students was used to model their temporal learning behaviours. After pre-processing data, feature engineering, sequence construction, and normalisation, two RNN models were employed and evaluated based on accuracy, precision, recall, F1-score and ROC-AUC performance metrics. The classification accuracy for both models was around 60%, with similar results. Their predictive performance, though, was still modest, with ROC-AUC around 0.50 and recall less than 25%, which means that there was not much discriminatory power between at-risk and retained students. The feature importance analysis showed that the interactions in LMS were weak predictors of dropout, meaning that the information from behavioural engagement alone is limited in predicting dropouts. The results highlight the need for the development of more powerful predictive models using multimodal ensembles which combine the LMS interaction sequences with academic, demographic, and behavioural data. This study offers empirical proof of the challenges of sequence-based RNNs for educational risk prediction and lays the groundwork for building more powerful, explainable and intelligent early warning systems for timely intervention and student retention in digital learning environments.