Balancing early detection and performance in predicting students at risk of dropout on MOOCs using time-series analysis and a hybrid deep learning model

Abstract

Student engagement functions as the main factor which determines how effective online educational programs will be. For this reason, Massive Open Online Course (MOOC) platforms must identify students who might withdraw the course at specific times to implement timely interventions. Current identification methods focus primarily on final accuracy metrics to identify their target students. The study proposes a Multi-Input Convolutional Bidirectional Long Short-Term Memory model (MI-CBiL), which integrates static enrollment-history features with dynamic daily interaction matrices for early MOOC dropout prediction. The model jointly processes static enrollment-history features and dynamic daily interaction matrices, while a weighted multi-objective criterion is subsequently applied to select a prediction window that balances predictive accuracy and intervention timeliness. The MI-CBiL model achieved 95.20% accuracy on the KDD Cup dataset while surpassing the baseline architectural results. An Early Warning configuration identified at-risk students within the initial days of enrollment with 93.11% accuracy, whereas a High Precision setup saved 35.9% of the course duration while achieving 95.34% accuracy. The framework delivers an effective solution which enables rapid MOOC retention improvements through its proactive approach.