Ensemble Learning for Early Academic Performance Prediction: A Case Study in Engineering Leveling Courses

Abstract

Student attrition in engineering leveling courses is a persistent challenge in Latin American higher education, often exacerbated by the lack of timely identification mechanisms. Traditional Early Warning Systems (EWS) frequently rely on mid-term grades, limiting their capacity for preventive intervention. This paper presents an Ensemble Learning model designed to predict final academic performance using exclusively pre-enrollment data, combining sociodemographic, previous academic records, and socioeconomic and psychometric variables. This study analyzed historical records of 1,533 students (7,210 instances) from a polytechnic university in Ecuador. Five machine learning algorithm families were evaluated through rigorous hyperparameter optimization involving 9,972 configurations. The proposed Stacking architecture, which integrates the top-performing XGBoost models via a Ridge Regression meta-learner, achieved a Coefficient of Determination (R2) of 0.749 and a Mean Absolute Error (MAE) of 3.783 on a 0-40 scale, demonstrating superior generalization compared to single models. Furthermore, SHAP (SHapley Additive exPlanations) analysis revealed that non-cognitive factors, such as study habits and academic self-confidence, are critical predictors alongside high school GPA and mathematics admission scores. To validate its practical utility, the model was deployed for the 2025-B cohort (N=900), identifying 9.2% of incoming students as "High Risk" before the start of classes. These results demonstrate the technical feasibility and ethical viability of using advanced ensemble methods for proactive, evidence-based academic management.