<![CDATA[Student attrition poses a substantial challenge to higher education institutions, affecting their reputation and financial sustainability. Conventional single machine learning models often exhibit limited sensitivity when analyzing educational data, which is typically marked by severe class imbalance favoring graduating students over dropouts. This study introduces an Early Warning System based on a Hybrid Stacking Ensemble framework to improve student attrition prediction. The approach leverages complementary biases from Bagging and Boosting as base learners, which are then combined using a Logistic Regression meta-learner to refine prediction weights. To counteract class imbalance and majority-class bias, the Synthetic Minority Over-sampling Technique was employed during preprocessing. Empirical evaluations reveal that the Hybrid Stacking Ensemble attains a classification accuracy of 88.81% and a Recall of 80.99%, surpassing standalone models and other ensemble methods. Feature importance rankings highlight second-semester academic performance and administrative-financial factors—particularly tuition payment punctuality—as key dropout predictors. These results affirm the value of integrating diverse classifiers to discern intricate, nonlinear student behavior patterns. In essence, this work establishes a reliable, evidence-based framework enabling administrators to shift from reactive to proactive, precision-targeted strategies that foster student retention and institutional success.]]>
Copyrights © 2026
