Aplicação de técnicas de mineração de dados e machine learning para predição da evasão de estudantes nos cursos de graduação de uma IFES

Abstract

Student dropout in Brazilian higher education is a phenomenon of considerable social and economic relevance, as it represents a waste of public resources, disrupts human capital development, and exacerbates structural inequalities. Within Brazilian Federal Institutions of Higher Education (Instituições Federais de Ensino Superior – IFES), this problem is intensified by the absence, in many institutions, of systematic mechanisms for the early identification of students at risk of dropping out. The Federal University of Fronteira Sul (Universidade Federal da Fronteira Sul – UFFS), whose Undergraduate Success Rate was approximately 33.11% between 2015 and 2023, does not have a consolidated institutional predictive model for this purpose. In this context, this study aimed to develop and evaluate predictive models of student dropout in undergraduate programs at the UFFS Chapecó Campus using data mining and machine learning techniques. Methodologically, this is an applied, quantitative, and descriptive study conducted as a case study. The analytical process followed the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology. Data were obtained from four institutional information systems—SIGAA, SCI, SAS, and Pergamum—resulting in an analytical dataset comprising 235,632 records related to 8,749 student enrollments from the first semester of 2014 to the second semester of 2024. Six supervised machine learning algorithms were trained and compared: Logistic Regression, Decision Tree, Random Forest, XGBoost, Linear Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). Validation was performed using a five-fold StratifiedGroupKFold protocol, grouped by student enrollment, ensuring that records associated with the same enrollment were included exclusively in either the training set or the test set within each fold. This procedure reduced the risk of information leakage between records associated with the same student. The XGBoost model achieved the best performance, with an AUC-ROC of 0.8878, accuracy of 0.8039, and macro-averaged F1 score of 0.8037, followed by Logistic Regression (AUC-ROC = 0.8681) and Linear SVM (AUC-ROC = 0.8680). The feature importance analysis revealed that academic performance predominated in the model’s predictive structure: students’ final course grades accounted for approximately one-third of the total information gain, followed by the number of absences and field of study. The use of the University Restaurant during the initial semesters emerged as a predictor of institutional engagement, surpassing sociodemographic variables such as gender and ethnicity. It is concluded that the developed models demonstrated satisfactory predictive capacity and institutional applicability. As a technical and technological product, this study presents a technical report that systematizes the experience of applying predictive models within the context of the UFFS Chapecó Campus and proposes guidelines for implementing an early warning system integrated with the institution’s information systems. The proposed approach has the potential to be replicated in other Brazilian Federal Institutions of Higher Education, contributing to the development of evidence-based student retention policies.