A Feasibility-Aware Counterfactual Framework for Explainable Early Academic Risk Assessment: A Synthetic-Data Prototype

Abstract: Early academic risk prediction can support timely educational intervention, but a risk score alone does not explain why a student is flagged or what realistic changes may reduce that risk. This paper presents a feasibility-aware counterfactual framework that combines supervised risk prediction, local explainability, constrained counterfactual generation, and action-oriented reporting. A fully synthetic dataset of 2,500 student records was generated for controlled methodology validation. Four classifiers were evaluated using a stratified 75:25 hold-out split. Logistic Regression achieved the highest F1-score (0.567) and ROC-AUC (0.805) in this prototype. SHAP was used to explain model behaviour, while the counterfactual module searched bounded changes in potentially actionable academic- behavioural variables. Valid counterfactuals were found for 70.8% of the evaluated high-risk cases, with a mean of 1.86 changed features. The study is explicitly a synthetic-data prototype; its results demonstrate methodological feasibility and reproducibility rather than effectiveness on real students.

Keywords: Academic Risk Prediction, Explainable Artificial Intelligence (XAI), Counterfactual Explanations, Educational Data Mining, Early Warning Systems, SHAP, Actionable Recourse, Machine Learning.