Student Dropout Risk Prediction After Two Semesters: A Comparative Study Of Logistic Regression, Random Forest, and XGBoost with SHAP

Abstract

Student dropout requires accurate and explainable early detection so that model outputs can support academic decisions responsibly within an early warning system. This study compares logistic regression, Random Forest, and extreme gradient boosting for student dropout prediction and examines how explainable artificial intelligence clarifies influential predictive factors through Shapley additive explanations. The Predict Students’ Dropout and Academic Success dataset from the Machine Learning Repository contains 4,424 observations and 36 features. The Enrolled class was excluded because it is not a final outcome, producing a binary dataset of 3,630 observations comprising 1,421 Dropout and 2,209 Graduate records. Data were stratified into 60% training, 20% validation, and 20% test sets. Hyperparameters and classification thresholds were selected on validation data, while the test set was reserved for final evaluation. Logistic regression achieved the highest precision-recall area under the curve of 0.9728 and Dropout recall of 0.9754. Random Forest produced the highest F1-score of 0.8346 and balanced accuracy of 0.8710. Extreme gradient boosting achieved a precision-recall area under the curve of 0.9705, recall of 0.9613, precision of 0.7319, F1-score of 0.8311, and balanced accuracy of 0.8675. Shapley additive explanations ranked study program, second-semester approved curricular units, tuition payment status, and early academic achievement among the dominant features. Extreme gradient boosting is competitive but not universally superior. Local validation, temporal evaluation, fairness assessment, and intervention testing are required before operational deployment.

GS Cited Analysis