Article

Scholarship Prediction for International Students Using Machine Learning: A Temporal Stability Analysis Across Enrollment Years

---

Abstract

Scholarship allocation plays a critical role in international higher education by shaping access, mobility, and institutional strategy. This study investigates whether scholarship receipt can be predicted using observable migration and academic descriptors from a global student mobility dataset spanning 2019–2023. The prediction task is formulated as a binary classification problem with scholarship status as the target variable. To ensure realistic assessment, two temporally structured evaluation protocols are employed: leave-one-year-out validation and forward-chaining validation. Multiple model families are compared, including Logistic Regression, Random Forest, XGBoost, LightGBM, and CatBoost. Performance is evaluated using discrimination metrics (ROC-AUC and PR-AUC), threshold-based metrics (F1, balanced accuracy, and Matthews Correlation Coefficient), and calibration via Brier score. Feature relevance is examined using permutation importance. Across all models and temporal splits, predictive performance remains close to random classification. Mean ROC-AUC values range between approximately 0.50 and 0.52, balanced accuracy remains near 0.50, and MCC values cluster around zero. Brier scores approximate the baseline value expected under uniform probability predictions. Permutation importance magnitudes are small and unstable across model types, indicating minimal and inconsistent feature contribution. These findings collectively suggest that scholarship receipt is effectively unpredictable using the available migration and academic attributes. The results highlight the limits of applying machine learning to allocation decisions when core determinants—such as academic merit, financial need, and institutional policy criteria—are absent from the dataset. By emphasizing temporal validation and multi-metric evaluation, this study demonstrates the importance of rigorous methodological design in educational analytics. The findings contribute to responsible AI discussions by illustrating how weak predictive signal can persist despite advanced modeling techniques, underscoring the necessity of comprehensive and decision-relevant data for meaningful scholarship prediction.

Keywords: Scholarship Prediction, Temporal Validation, Educational Data Mining, Permutation Importance, Responsible AI

How to Cite:

Hery, H., (2026) “Scholarship Prediction for International Students Using Machine Learning: A Temporal Stability Analysis Across Enrollment Years”, Artificial Intelligence in Learning 2(1). doi: https://doi.org//AIL.143