Predicting Student Motivation and Engagement Using a Hybrid of Recurrent Neural Networks (RNNS) and Reinforcement Learning

Abstract

Student motivation and engagement are pivotal determinants of academic achievement, yet their Dynamic and evolving characteristics create difficulties for conventional predictive analytics models. In this paper, a new hybrid framework is introduced that leverages recurrent neural networks based on long short-term memory (LSTM) and deep Q-network (DQN) reinforcement learning to anticipate and optimize student motivation and engagement in intelligent tutoring systems (ITS). The RRHEP model is trained based on sequential data of learning interactions, such as clickstream events, quiz scores, length of sessions, forums engagement, and sentiment-tagged feedback messages. Its performance is validated using the OULAD dataset, comprising 32,593 users and 22 instances of courses. According to the experimental results, RRHEP achieves an accuracy of 94.7%, an F1-score of 0.932, an AUC-ROC of 0.971, an MAE of 0.038, and an RMSE of 0.051, outperforming five baseline methods, including logistic regression, support vector machines, regular LSTM, bidirectional LSTM, and DQN alone. In addition, ablation studies reveal that the temporal ordering modeling unit and the reinforcement-driven feedback policy are indispensable components for the RRHEP model, leading to improvements by 8.3% and 6.1%, respectively. Moreover, the RRHEP intervention model exhibits a significantly shorter reaction time by 42% compared to the feedback policy. Based on the above results, RRHEP is demonstrated to be an effective and scalable model in predicting user engagement for personalized learning platforms. Future research could explore other promising techniques, such as multimodal fusion and federated learning.