Predição de evasão estudantil usando aprendizado de máquina

Abstract

In recent years, dropout rates in Higher Education have intensified, causing negative impacts on students and institutions, particularly in the public sector, which suffers significant financial losses. Various factors contribute to this phenomenon, such as learning difficulties, inadequate course structures, and a lack of financial resources. This study proposes using Machine Learning (ML) algorithms to predict and identify behavioral patterns of students at risk of dropping out. For this purpose, experiments were conducted using data extracted from the academic system of the Federal Technological University of Paraná (UTFPR), Dois Vizinhos campus. The data were properly anonymized and pre-processed to ensure that sensitive information could not be accessed by the predictive models. Experiments were conducted with different predictive tasks, exploring the completeness of the records extracted from the university’s system. Additionally, four "interpretable" algorithms were explored, which provide rules or structures that allow managers to understand predictions and make informed management decisions. The experimental results showed that the best-performing algorithm was Random Forest, achieving the highest average F1-Score across all experiments, with a maximum value of 0.99. This result was obtained using the dataset containing only the latest student records from the academic system, considering a binary classification between “Regular vs. Dropout“ students, with attribute thresholds of 0.05 for constants and 0.8 for correlations. Although the results are promising, they still lack the inclusion of socioeconomic variables, which limits the system’s ability to understand the causes of dropout fully