Predicting Student Dropout in Higher Education Using Explainable Machine Learning: Evidence from Bangladesh

Abstract

Student dropout in higher education is a critical institutional challenge in Bangladesh, where over 249,000 university students discontinued their studies between 2020 and 2021. This paper presents an interpretable machine learning framework for dropout prediction using a multidimensional dataset of 891 students from Khwaja Yunus Ali University (KYAU), integrating academic, socioeconomic, and psychological variables. Conditional Tabular Generative Adversarial Network (CTGAN) mitigates data scarcity within a leakage free pipeline benchmarking six classifiers under stratified cross validation. Random Forest attains the best composite score, with test ROC-AUC of 0.9923 and F1-score of 0.9637. SHapley Additive exPlanations (SHAP) identifies depression score as the dominant dropout predictor, surpassing conventional academic indicators; performance remains strong (ROC-AUC 0.9818) with this feature removed. A sensitivity analysis showed statistically comparable performance without CTGAN augmentation, indicating that augmentation did not inflate the reported results. These findings demonstrate a robust and interpretable prediction framework suitable for supporting institutional early warning interventions. An advisor facing Gradio interface translates per student SHAP explanations into actionable intervention signals. External validation at Varendra University, using an independent sample of 187 students, yielded ROC-AUC of 0.7807, confirming genuine cross-institutional signal and motivating site-specific recalibration before deployment while demonstrating practical potential for scalable early warning across universities.