FAIRNESS-AWARE EXPLAINABLE AI FOR EARLY IDENTIFICATION OF UNIVERSITY STUDENT DROPOUT RISK IN PAKISTAN

Abstract

University student dropout remains a significant challenge for higher-education institutions, particularly where timely identification of academically and socioeconomically vulnerable students is limited. This study developed a fairness-aware explainable artificial intelligence (AI) framework for early prediction of university student dropout risk in Pakistan. A quantitative retrospective predictive design was employed using student academic, attendance, behavioral, engagement, socioeconomic, and institutional data. Logistic Regression, Random Forest, and XGBoost models were developed and comparatively evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and PR-AUC. Explainable AI techniques, particularly SHAP, were applied to identify the principal factors contributing to individual predictions, while fairness metrics were used to examine disparities across student groups. The results indicated that XGBoost achieved the strongest predictive performance (accuracy = 87%, ROC-AUC = .92), with GPA, failed courses, attendance, LMS engagement, and financial difficulty emerging as major predictors. Fairness analysis indicated disparities in prediction outcomes across student groups; however, fairness mitigation substantially reduced these disparities with only a modest reduction in predictive performance. The findings demonstrate that integrating predictive accuracy, explainability, and fairness can provide a more transparent and equitable approach to student dropout-risk identification. The proposed framework offers a basis for responsible AI-supported early-warning systems and targeted student-support interventions in Pakistani higher education.