Modelagem preditiva explicável para apoio à prevenção da evasão em cursos de graduação

Abstract

Academic dropout remains a recurring challenge in higher education institutions, generating personal impacts for students as well as relevant social and institutional consequences. In recent years, machine learning models have been used to estimate the risk of academic dropout based on students’ academic, demographic, and socioeconomic data. However, despite their good predictive performance, many of these models operate with limited transparency, making it difficult to interpret their results. In this context, Explainable Artificial Intelligence (XAI) approaches emerge as an alternative to make predictions more understandable and practically useful. This dissertation investigates how Explainable Artificial Intelligence techniques can be applied to predictive models to estimate the risk of academic dropout in undergraduate programs, aiming to identify the factors that most influence model decisions and to analyze the potential of the generated explanations as support for planning institutional strategies to prevent dropout. The research was conducted using data from on-campus undergraduate programs at the Federal University of Maranhão (UFMA) and adopted a structured analytical approach that combines exploratory analysis, course-segmented predictive modeling, and global and local explainability techniques. Different machine learning algorithms were used to estimate dropout risk, and XAI methods such as SHAP, Permutation Importance, LIME, and Anchors were applied to interpret predictions and analyze the contribution of variables to model decisions. The results indicate that predictive models show good performance in identifying students at risk, although with variations across programs. The explanatory analyses revealed convergence among techniques in identifying three academic dimensions as central to the structure of predictions: efficiency in converting enrolled workload into course approvals (IECH), the pace of progression throughout the program (IEPL), and overall academic performance represented by the completion GPA (MC). In the analyzed context, demographic variables showed less significant influence on predictions. Furthermore, the results suggest that integrating predictive modeling with explainability techniques makes dropout risk more interpretable by highlighting academic performance patterns associated with the event. The explanations help clarify which factors sustain risk estimates across different programs. Consequently, this approach can support academic management in identifying student profiles that require greater attention and in planning institutional actions aimed at promoting student retention.