A Machine Learning-Based Student Dropout Risk Prediction and Early Warning System

Abstract

Student dropout rates in higher education have become a significant problem for universities worldwide, impacting both student academic achievement and institutional stability. To address this problem, this research proposes the Student Dropout Risk Prediction and Early Warning System (SDRP-EWS), a machine learning framework designed to predict the risk of eventual student dropout. Utilizing multi-semester academic performance data from 20,919 undergraduate students at Sakon Nakhon Rajabhat University, Thailand (2014–2024), the study developed logistic regression, random forest, and XGBoost models. To support early and continuous monitoring, these models were evaluated at three distinct prediction horizons: following the first semester, after the second semester, and after three or more semesters. Cohen's d effect size analysis compared ROC-AUC distributions across engineered feature sets, confirming the practical advantage of the selected model configurations. The results show that XGBoost consistently outperformed other models across all prediction stages, achieving F1-scores for the dropout class ranging from 0.65 in the early stages up to 0.93 in later semesters with comprehensive academic histories. This study's primary contribution is integrating the top-performing multi-semester model into a functional, web-based prototype dashboard tailored for Thai higher education. By providing academic advisors with intuitive risk scores and key contributing factors, the system demonstrates strong potential as an evidence-based decision-support tool to facilitate proactive, targeted institutional interventions.