<![CDATA[Predicting students at risk of delayed graduation is essential for enabling timely academic intervention, yet educational datasets are often characterized by class imbalance that limits predictive performance. This study proposes and evaluates an optimized XGBoost framework that integrates SMOTE-ENN, hyperparameter tuning, and decision threshold adjustment for student graduation prediction. A quantitative machine learning approach was conducted using academic records from 315 alumni across multiple Indonesian universities. Six classification algorithms were systematically compared to identify the most suitable baseline model before optimization. Model performance was assessed using multiple classification metrics to ensure comprehensive evaluation. The findings demonstrate that XGBoost consistently outperformed the competing algorithms and achieved its strongest predictive performance after integrating all three optimization strategies. Compared with applying each optimization technique individually, the combined framework produced more balanced classification results, improved minority-class recognition, and reduced prediction bias caused by imbalanced data. Feature analysis further revealed that academic variables, particularly cumulative grade point average, accumulated credits, and course repetition history, were the strongest predictors of timely graduation, whereas social and non-academic variables contributed comparatively less. These findings provide an effective and replicable machine learning framework for early identification of students at risk of delayed graduation and offer practical support for data-driven academic intervention and decision-making in higher education]]>
Copyrights © 2026
