Comparative Analysis of Machine Learning Algorithms for Academic Performance Prediction Using Oversampling Techniques

Predicting academic performance has become increasingly critical for educational institutions seeking to implement proactive student support strategies and improve retention rates through early intervention programs.While machine learning approaches demonstrate considerable promise for identifying at-risk students, the impact of class imbalance correction methods and algorithm selection on prediction accuracy remains inadequately understood, particularly within Latin American higher education contexts where institutional and cultural factors may influence the applicability of predictive models developed in other regions.This study systematically evaluates the effectiveness of six machine learning algorithms-Decision Tree, Support Vector Machine, Bayesian Networks, K-Nearest Neighbors, Logistic Regression, and XGBoost-combined with oversampling techniques to predict multi-class academic performance in a Peruvian public university.From an initial dataset of 4,584 student records and 26 attributes, data cleaning reduced it to 2,656 records and 11 features.Class imbalance was addressed using oversampling to ensure equal representation across the academic performance categories.Among the tested algorithms, XGBoost achieved the highest performance with 98.10% accuracy, 98.09% precision, 98.08% recall, and a 98.08% F1-score.These results suggest that ensemble methods, particularly XGBoost, offer superior predictive capability when paired with data balancing techniques.This study is limited to a single-institution dataset; future research should validate the findings using multi-institutional data.