PREDICTING ACADEMIC SUCCESS WITH MACHINE LEARNING: A SIMULATION-BASED ANALYTICS FRAMEWORK FOR A REGIONAL UNIVERSITY IN UZBEKISTAN

Abstract

Higher education institutions in developing regions increasingly generate large volumes of student data, yet most continue to rely on manual, retrospective evaluation rather than predictive analytics. This study examines whether supervised machine learning can support early identification of academically at-risk students at Termez State University, Uzbekistan, a context in which empirical evidence on educational analytics remains scarce. A simulated institutional dataset of 2,450 undergraduate records, constructed to reflect realistic academic, attendance, demographic, and digital-engagement parameters, was used to train and compare four classifiers: Logistic Regression, Support Vector Machine, Random Forest, and Artificial Neural Network. Academic success was operationalised as a binary outcome (GPA ≥ 2.50 versus GPA < 2.50). Descriptive and correlation analyses preceded model training on an 80/20 split, with performance evaluated via accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. Random Forest achieved the strongest performance (accuracy = 0.89, F1 = 0.86, ROC-AUC = 0.91), outperforming the Artificial Neural Network (0.87), Support Vector Machine (0.85), and Logistic Regression (0.82). Final examination score, attendance rate, assignment score, and previous GPA were the most influential predictors, while demographic variables contributed minimally. Approximately 24% of students were classified as academically at risk. The findings suggest that ensemble learning methods offer a feasible, interpretable foundation for early-warning systems in resource-constrained regional universities, while underscoring the need for real institutional data, broader psychosocial variables, and explicit governance safeguards before operational deployment.