Optuna-Optimized Meta-Learner and Ensemble Learning Models for Student Performance Prediction

Assessing and predicting students’ academic performance remains a significant challenge in applying machine learning to education. This study explores the effectiveness of ensemble learning methods in improving predictive accuracy. Using an academic dataset from a public university, we evaluated multiple individual models alongside ensemble techniques, including hard voting, soft voting, and stacking. To optimize hyperparameters and enhance model robustness, we employed three optimization methods: GridSearchCV, HalvingGridSearchCV, and Optuna. Among these, Optuna demonstrated superior efficiency in hyperparameter tuning. Model performance was assessed using rigorous metrics, including accuracy, F1-score, AUC-ROC, kappa, and sensitivity. Our findings highlight the superiority of the stacking model, which hierarchically combines predictions from multiple algorithms, outperforming traditional voting-based approaches. This improvement is attributed to the complementary nature of the base models and the fine-tuned hyperparameter optimization enabled by Optuna. To enhance decision interpretability, we integrated explainable AI (XAI) techniques, particularly SHAP (SHapley Additive exPlanations). This approach identifies the most influential variables in the model’s predictions, ensuring transparency and providing deeper insights into the factors affecting academic success.