An Explainable Feature-Engineered Stacking Ensemble Framework for Student Outcome Prediction Using Learning Analytics

Abstract

Making accurate predictions about student academic outcomes is critical to ensure that interventions can be made in time to enhance student learning in the contemporary learning environment. But, a lot of the current prediction models focus mainly on prediction accuracy without offering much interpretability and that makes them hardly useful when it comes to educational decision making in practice. In this paper, we present an Explainable Feature-Engineered Stacking Ensemble Framework for predicting students' outcomes based on the data from Open University Learning Analytics (OULAD). The proposed system consists of data preprocessing, feature engineering, MIFS, stacking ensemble learning, and SHapley Additive exPlanations (SHAP) which is used to build an accurate and interpretable prediction model. A total of seven machine learning models (Random Forest, Extra Trees, XGBoost, LightGBM, CatBoost, Support Vector Machine, Logistic Regression) were initially tested, and the three most accurate models (LightGBM, CatBoost and XGBoost) were subsequently stacked together using a Logistic Regression as a meta-classifier to give the proposed stacking ensemble model. Results of the experiments showed that the proposed framework successfully achieved a classification accuracy of 87.62%, Precision of 85.34%, Recall of 89.08%, F1-score of 87.17% and a ROC-AUC of $0.9494$. Moreover, the mean accuracy of 10-fold cross validation was 87.18%, which is stable with a standard deviation of 0.60%, demonstrating good generalization results. The top four features that most affected student performance were found to be total assessments, average score, score standard deviation and the number of active learning days through SHAP-based explainability analysis. The suggested framework is robust and open, and can help education institutions to identify at-risk learners and to take appropriate academic interventions in time.