Predictive Modeling for Student Dropout in Online Courses Using Ensemble Deep Reinforcement Learning

Abstract

The increase in students quitting online educational platforms is a critical problem that can potentially jeopardize education, student growth, and organizational effectiveness. This paper proposes the use of the EDRL method for the prediction of student dropout in Massive Open Online Courses (MOOCs) as well as other online educational platforms. The proposed algorithm entails three distinct techniques, including (i) application of LSTM networks for pattern extraction in sequence; (ii) use of DDQN for feature set selection; and (iii) use of a gradient boosting ensemble of several predictive models. The EDRL algorithm was evaluated using the Open University Learning Analytics Dataset, which involved data of 32,593 students in 32593 rows and 22 attributes related to 7 modules. The experimental assessment shows that the EDRL model attains an accuracy score of 94.7%, F1 Score of 93.9%, Area Under the Curve-ROC (AUC-ROC) of 0.971, precision of 94.2%, and recall of 93.7%, outperforming seven different models, including standalone LSTM, XGBoost, Random Forest, SVM, CNN-LSTM, traditional DQN, and Logistic Regression. Furthermore, an ablation study confirms the effectiveness of individual components of the architecture in enhancing its overall performance. The feature dependency of the RL agent is reduced by 31.4% while simultaneously improving the generalization rate for unseen students by 8.6%. It can thus be concluded that the presented model offers a highly promising approach to identifying students at risk and thereby taking early actions for improvement in learning outcomes.