Predicting Student Dropout Risk in Online Learning Using Stacked Ensemble Machine Learning and Explainable AI Techniques

Abstract

Predicting student dropout in online learning platforms such as MOOCs and institutional LMS platforms, is a critical challenge in educational data mining. Although numerous machine learning models have been proposed to predict dropout likelihood, the lack of model interpretability has limited their practical deployment in educational settings. This paper proposes a stacked ensemble machine learning model combining Logistic Regression, Random Forest, and XGBoost, with explainable AI techniques to identify at-risk learners using behavioral and demographic features. The dataset, obtained from Kaggle's MOOC Dropout Prediction challenge, was cleaned, balanced, and subjected to feature selection to prevent information leakage. With SHAP interpretability, the model achieves an accuracy of 65%, ROC AUC of 0.71, and PR AUC of 0.73. Our results show that dropout prediction is feasible using early behavioral data, and stacked models offer a promising balance of performance and transparency. This work contributes a replicable, explainable architecture suitable for real-time educational intervention systems.