Ex-ADA: a SHAP-based explainable AdaBoost framework for predicting at-risk students

Introduction Early identification of academically at-risk students remains a persistent challenge in higher education, largely due to the limited explainability and adaptability of existing predictive models. Although many early-warning systems rely on behavioral, assessment, or attendance data, their lack of transparent decision-making often reduces trust and limits their practical utility for educators. Methods To address this problem, this study proposes Ex-ADA, an Explainable AdaBoost-based framework that integrates the interpretive strength of SHapley Additive exPlanations (SHAP) with the robust ensemble learning capabilities of AdaBoost. Using academic, behavioral, and engagement indicators from 642 students enrolled in the Fundamentals of Programming course at the University of Plovdiv, the framework aims to deliver both high predictive accuracy and human-interpretable insights for data-driven intervention. Results Ex-ADA achieves an accuracy of 84.12% and an AUC of 92.31%, outperforming conventional classifiers such as k-nearest neighbor, decision tree, naïve Bayes, and multilayer perceptron. SHAP analyses reveal that attendance, midterm practice performance, and homework completion are the most influential predictors of student success. Discussion In addition to global interpretability, the framework provides personalized, instance-level explanations that help instructors understand each student’s risk factors. By bridging predictive analytics with transparent educational decision-making, Ex-ADA demonstrates how explainable ensemble models can enhance early-warning systems and support more effective, timely pedagogical interventions.