Mineração de dados educacionais para identificação do risco de evasão: uma análise comparativa de algoritmos de classificação no curso de Sistemas de Informação da UFU

Abstract

Academic dropout is a relevant challenge for higher education institutions, especially in Computer Science programs. This study aims to comparatively evaluate the performance of supervised classification algorithms in identifying an operational indicator of dropout risk among students enrolled in the Bachelor’s Degree in Information Systems at the Federal University of Uberlândia. A dataset comprising student records from the School of Computer Science was used; these records were originally organized by individual courses taken. After cleaning, transformation, and aggregation, each record represented a student’s performance by academic year and semester, considering the number of courses taken, courses passed, courses failed, and the academic performance coefficient. The class variable evadiu was assigned the value 1 when the number of failed courses exceeded the number of passed courses and 0 otherwise; therefore, it represents a risk indicator rather than institutionally confirmed dropout. The Naive Bayes, J48, Random Forest, IBk, Logistic, and JRip algorithms were evaluated with two parameter settings each using stratified 10-fold cross-validation. Logistic achieved 100% accuracy in both settings; Random Forest, JRip, and J48 exceeded 99.9%; IBk achieved performance close to 100%; and Naive Bayes obtained accuracy between 95.7% and 97.0%. However, because the class was constructed from attributes that were also used as predictors, the high scores reflect the strong relationship between this rule and the data representation. Thus, the comparison identifies Logistic as the best-performing classifier under the investigated setup, but it does not demonstrate an ability to predict definitive institutional dropout.