From prediction to prescription: profile-aware prescriptive analytics for student dropout in a Latin American Public University

Abstract

Student dropout in higher education remains a persistent challenge, particularly in developing contexts where structural disadvantage intersects with academic readiness. This study proposes a prescriptive analytics framework that integrates supervised learning, unsupervised profiling, and explainable artificial intelligence to move beyond prediction toward actionable interventions. Using anonymized records from 2,500 entrants (2020–2024) at a Latin American public university, the analysis combines data from a foundation (pre-university) program with early bachelor's indicators and socioeconomic profiles. A Random Forest classifier was evaluated under stratified cross-validation, while student profiling and explainable AI techniques (SHAP, ALE, DiCE, CARLA, and Alibi) were used to characterize risk patterns and generate actionable recommendations. The Random Forest classifier achieved strong discrimination, with ROC-AUC values between 0.95 and 0.97 and an Expected Calibration Error of 0.031. Student profiling revealed three coherent segments with observed dropout ranging from less than one percent to approximately one quarter. Global and local explainability consistently identified foundation GPA, early credit progress, socioeconomic index, and residential Internet access as the primary drivers of predicted dropout. Counterfactual explanations identified minimal, feasible changes that reclassified high-risk cases as lower predicted risk, providing recourse over the model's decision. The proposed framework extends predictive modeling by generating interpretable, profile-aware, and actionable recommendations that can support equitable student success through targeted interventions.