Predicting At-Risk Students from Administrative Data Using Machine Learning: A Multi-Factor Analysis at a Cambodian University

Abstract

Early identification of students who fail to meet minimum academic standing allows universities to intervene before academic damage accumulates. This study evaluates whether routinely collected administrative records alone can identify such students at Paragon International University, Phnom Penh, Cambodia. The dataset covers 3,650 students with a recorded cumulative GPA (CGPA) from 15 departments in five faculties across 16 academic years (2010–2026), and includes currently enrolled, graduated, and withdrawn students. A student was labeled at-risk when their CGPA fell below 2.0, the university’s minimum passing standard. Eight administrative features (gender, age, department, enrollment year, credits earned, scholarship status, scholarship amount, and grade level) were used to train four classifiers: Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting. Gradient Boosting performed best on the held-out test set (n = 730), with 89.18% accuracy, an F1-score of 0.675, and an AUC-ROC of 0.928. Credits earned was the dominant feature (42.46% importance), followed by age (12.88%) and enrollment year (11.97%). Overall, 17.6% of students were at-risk, and scholarship recipients had a higher mean CGPA (2.76) than non-recipients (2.23). Because several features accumulate during enrollment, the models classify current academic standing rather than forecast it at entry; prospective early prediction is identified as the next step.