Leakage-Safe Ensemble Learning for Early Prediction of Student Degree Outcomes in a Nigerian University

This paper proposes a leakage-controlled ensemble learning framework for the staged prediction of undergraduates' final degree class from longitudinal student records of a Nigerian university.The application context is early warning/advising systems where there is a need to identify students suitable for interventions prior to graduation.Anonymous academic history data of students who graduated from a five-year Computer Science and Engineering program were prepared into three-, four-and five-year prediction points.Potential predictor attributes comprised students' year-wise cumulative grade point average (CGPA), average scores, cumulative registered units and academic standing.Leakage was controlled by restricting information available for early predictions to exclude students' final CGPA, final degree class, and attributes from future academic years.Random Forest (RF) and Extreme Gradient Boosting (XGBoost) classifiers were evaluated against each other under a validation framework designed for both information leakage and class imbalance.Metrics for evaluation include accuracy, macro-and weighted-F1, balanced accuracy and confusion matrix error distributions instead of accuracy alone.Preliminary findings indicate that overall, XGBoost yielded better and more consistent classification performances than RF especially at the beginning and end of the academic program, while RF had comparable performances in the intermediate years.Ensemble learning methods could be utilized for data-informed student advising if prediction outcomes are used as intervention signals instead of categorical academic decisions.We share insights from our staged prediction framework, leakage control heuristics and an in-house deployment path for educational data mining in developing world universities.