Explainable Machine Learning Pipeline for Early Warning of Student Dropout in Graduate Programs Using SHAP and Gradient Boosting

Abstract

Student attrition in graduate programs is costly for students and institutions alike, yet most predictive models deployed for early warning behave as black boxes that program managers cannot audit. This paper proposes an end-to-end explainable machine learning pipeline for the early detection of dropout risk in master’s degree programs. The pipeline integrates academic, engagement, and administrative features available at the end of the first semester; trains a gradient boosting classifier with cross validated hyperparameter selection against logistic regression and random forest baselines; calibrates a cost sensitive alert threshold; and attaches Shapley additive explanations (SHAP) to every alert at both global and individual levels. The pipeline is evaluated in a controlled simulation study of 1,200 synthetic student records whose generative process mimics graduate-program registrar and learning management system data, including nonlinear threshold and interaction effects. On a held-out test set, the tuned gradient boosting model attains an area under the ROC curve of 0.790 with a recall of 0.613 at the calibrated threshold, comparable to the strongest baseline, while providing exact, efficient TreeSHAP explanations. Global SHAP analysis correctly recovers first-semester GPA, attendance, and engagement as the dominant risk drivers embedded in the simulation, and local explanations translate individual alerts into actionable counseling points. The results indicate that explanation quality, rather than raw discrimination alone, is the decisive ingredient for adoption of dropout early warning systems by study program managers. Keywords: Dropout pred