Student dropout prediction has become a critical field of research in learning management systems with the rapid growth of large-scale online education systems. Nevertheless, most available research is primarily dedicated to enhancing predictive accuracy and offers few mechanisms for converting predictions into effective interventions that help decrease student dropout rates. This is a major limitation, since high dropout rates negatively affect learner success, institutional efficiency, and the sustainability of online learning environments. To overcome this limitation, this research paper proposes a new student dropout prediction framework that uses deep learning networks together with an adaptive AI-based intervention system. The framework was evaluated using a large-scale HarvardX edX dataset containing more than 640,000 learner course interaction records. Three deep learning architectures, namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer, were developed and compared systematically. The experimental findings show that the Transformer model is the most effective, with a predictive accuracy of 0.92 and an AUC-ROC of 0.95, making it one of the most effective models for capturing temporal dependencies and behavioral patterns in normalized sequences of learner activities. Based on the prediction results, the individualized intervention program provided specific interventions, including motivational messages, learning materials, instructor notifications, and peer-support invitations. The use of the intervention led to improvements in learner engagement and retention by 21.8% and 17.1%, respectively. Moreover, the interpretability analysis showed that active learning days and participation in course events are the strongest predictors of student dropout.
Student dropout prediction has become a critical field of research in learning management systems with the rapid growth of large-scale online education systems. Nevertheless, most available research is primarily dedicated to enhancing predictive accuracy and offers few mechanisms for converting predictions into effective interventions that help decrease student dropout rates. This is a major limitation, since high dropout rates negatively affect learner success, institutional efficiency, and the sustainability of online learning environments. To overcome this limitation, this research paper proposes a new student dropout prediction framework that uses deep learning networks together with an adaptive AI-based intervention system. The framework was evaluated using a large-scale HarvardX edX dataset containing more than 640,000 learner course interaction records. Three deep learning architectures, namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer, were developed and compared systematically. The experimental findings show that the Transformer model is the most effective, with a predictive accuracy of 0.92 and an AUC-ROC of 0.95, making it one of the most effective models for capturing temporal dependencies and behavioral patterns in normalized sequences of learner activities. Based on the prediction results, the individualized intervention program provided specific interventions, including motivational messages, learning materials, instructor notifications, and peer-support invitations. The use of the intervention led to improvements in learner engagement and retention by 21.8% and 17.1%, respectively. Moreover, the interpretability analysis showed that active learning days and participation in course events are the strongest predictors of student dropout.
