Optimizing student retention prediction in higher education: A comparative analysis of supervised and unsupervised machine learning models

Abstract

This research addresses the problem of student dropout in higher education in El Salvador, using machine learning models. To do so, a historical dataset from a private university was used, which includes more than 83,000 student records between the years 2019 and 2024. In the exploratory analysis phase, unsupervised techniques such as Principal Component Analysis (PCA), K-means, and hierarchical clustering were applied to segment students and identify profiles with similar demographic, academic, and economic characteristics. Subsequently, supervised predictive models were implemented, including SVM, KNN, Random Forest, AdaBoost, and XGBoost, evaluating their performance using metrics such as accuracy and AUC-ROC. The results indicated that Random Forest offers greater predictive capacity, while clustering techniques allowed identifying risk groups without the need for prior labels. It was observed that students enrolled in in-person classes, with low academic performance and self-funded programs, are more likely to drop out. The research concludes that the combination of exploratory analysis and predictive models facilitates the design of effective strategies to improve student retention. Furthermore, it highlights the value of institutional use of these tools to develop proactive interventions, especially those targeting first-year students or those with vulnerable academic and economic circumstances.