Abstract
Machine learning (ML) is an artificial intelligence tool that focuses on learning by generating models using established algorithms that represent a given dataset. It can be used as a predictive tool for students' academic performance (AP) at both undergraduate and postgraduate levels. A cross-sectional analysis was conducted using academic records of 922 postgraduate students admitted to the High Institute of Public Health, Alexandria University, Egypt, between 2020-2024. Data included 22 features spanning pre-enrollment metrics, academic performance, and demographic traits. Classification algorithms, and regression models were trained on 75% of the dataset, validated via 5-fold cross-validation. Performance metrics included accuracy, precision, recall, AUC for classification, and MAE, RMSE, and R² for regression. Regression models outperformed classification models in AP prediction, with Ensemble (Soft Voting) achieving the highest accuracy (74.25%), lowest MAE (0.3383), and RMSE (0.4316). Among classification models, Random Forest demonstrated superior accuracy (71.43%) and AUC (0.87). Numerical features like the number of failed courses showed the strongest negative correlation with AP (r = -0.37). Key predictors included bachelor's university, major, department, and pre-enrollment CGPA. Feature importance analysis highlighted failed courses as the top determinant, followed by institutional and academic background variables. Regression-based ML models, particularly Ensemble (Soft Voting), proved more effective than classification approaches for predicting nuanced variations in AP. These findings enable institutions to prioritize early interventions for at-risk students, and optimize resource allocation. However, moderate R² values (0.3832) underscore the need to integrate psychosocial and behavioral factors in future studies.