An Explainable Stacking Ensemble Learning Framework for Early Student Attrition Prediction and Risk Stratification Using Educational Data Mining

Abstract

Student attrition poses a significant challenge in higher education. It negatively impacts academic performance, reputation, and resource management. This paper presents an XAI-based stacking ensemble for early prediction of student attrition with EDM methods.
We used a publicly available dataset on Kaggle, which contained information about 4,424 undergraduate students and their personal characteristics, academic status and socioeconomic background.
Class imbalance is dealt with using a cost-sensitive learning strategy with the parameter of scale_pos_weight combined with stratified train-test splitting. The proposed stacking ensemble framework used the Random Forest and XGBoost as base models and Logistic Regression as a meta-learner. The selected threshold of classification for early prediction was 0.20. The proposed model performance is assessed with several performance metrics, including Accuracy, Precision, Recall, F1-score, ROC-AUC, PR-AUC, as well as calibration analysis.
Experimental results showed that the proposed stacking ensemble outperformed individual base models and offered the best prediction accuracy with a high Recall value in early dropout prediction.
The developed model interpretability by SHAP and LIME showed that student performance and financial situation played a significant role in predicting the risk of attrition at early stages of study. The predicted probabilities are divided into the categories of Low, Medium and High-Risk using percentiles, to facilitate interventions for the students at risk. In this context, we offered an interpretable and applicable artificial intelligence-driven early warning system in higher education.