Research on personalized student management systems via integrated ensemble learning and Z-score standardization

As the push toward "smart campuses" accelerates, multi-source student records have created a practical need for early-warning models that can support academic-risk prediction, learning-support recommendation, and counseling resource allocation rather than replace human judgment. This study focuses on a binary personalized-management task: identifying students who may require academic or behavioral intervention from de-identified campus behavioral and academic records. To address the heterogeneity of these records, we propose a leakage-controlled predictive framework that combines Z-score standardization with a Stacking ensemble learning strategy. Z-score parameters are estimated only from the training folds and then applied to validation or test data, thereby reducing scale bias while preserving relative deviations in features such as GPA, attendance, library use, campus-card consumption, dormitory access, and online-learning behavior. The ensemble integrates XGBoost, LightGBM, and CatBoost as base learners and uses Logistic Regression as the meta-learner. Experiments on approximately 12,000 de-identified student-semester records from the 2022-2023 academic year show that the proposed method outperforms single-model baselines across accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC, and RMSE. The framework is intended to provide interpretable decision support for student services under privacy protection and human oversight.