EducAAr: integrating ontologies, machine learning, and XAI for academic dropout analysis in higher education institutions

Abstract

Student dropout in higher education is a complex problem with academic, institutional, and social impacts, which makes the development of approaches capable of supporting its analysis relevant. In this context, this dissertation proposes EducAAr (Educational Analysis Architecture), an architecture that articulates an ontology-based canonical model, machine learning models, and explainability techniques to produce interpretable analyses of student dropout risk. The research was conducted using the Design Science Research approach across two development cycles. In the first cycle, an ontology was developed to integrate academic, social, and institutional data, enabling information organization, consistency verification, and queries about student trajectories. In the second cycle, the architecture was extended with machine learning models, probabilistic calibration, SHAP-based explainability, and a visualization panel for the results. The technical evaluation was conducted using anonymized undergraduate data from the Federal University of Juiz de Fora, encompassing 16 course offerings and 7,731 students. The results indicate that the architecture can organize heterogeneous educational data, generate dropout risk estimates, and present the main factors associated with these estimates in an interpretable manner. The generated explanations indicated a predominance of variables related to academic performance in the initial periods, although admission factors, student assistance, scholarships, quotas, and student profile also appeared in specific contexts. Thus, the contribution of this dissertation lies in the proposal and technical evaluation of an integrated architecture for dropout analysis, without replacing institutional interpretation and without treating the identified associations as causal relationships.