An interpretable feature-augmented stacked ensemble framework for student dropout prediction

Abstract

Student dropout prediction is regarded as one of the significant research fields in educational data mining. It aids the sustainability of institutions and early interventions in schools. This study proposes an Enhanced Stacked Dropout Predictor (eSDP), which is a feature-enhanced hybrid stacking framework. The proposed framework combines heterogeneous base learner predictions with important original student-level features during meta-learning. As level-0 learners, Gradient Boosting, Support Vector Machine, Random Forest, and Logistic Regression were adopted, whereas the meta-learner was Gradient Boosting. To compare the proposed feature-augmented stacking with the stacking strategy, another feedforward neural network (FNN)-based stacking ensemble was implemented as a comparative baseline. A synthetic educational dataset was used for the experimental assessment with 7,000 student records. The proposed eSDP framework achieved the highest overall predictive accuracy of all the models tested, with an F1-score of 0.928 and ROC-AUC of 0.987. The framework showed competitive results and Friedman and Nemenyi analyses were used for comparative ranking of the models evaluated. The framework’s predictive ability was also evaluated at the cross-dataset level by replicating the framework on an independent education dataset that showed a decrease in predictive ability compared to the main dataset. Furthermore, transparent feature-level and meta-learning decisions were obtained via the explanations of the global and local interpretability analyses of SHAP. The results also show that the eSDP framework achieves good predictive performance and high interpretability and comparative performance assessment for predicting student dropout and providing educational decision support.