Indicadores preditivos para análise das taxas de evasão: um estudo de caso no curso de bacharelado em Sistemas de Informação

Abstract

This study proposes indices for predicting student dropout rates in a course in the field of computing at a higher education institution, the Bachelor’s program Information Systems (BSI) at the Federal University of Rio Grande do Norte (UFRN), using machine learning (ML) and dimensionality reduction. To this end, four indices were proposed as promising predictors of student dropout: persistence, failure rate, semester-by-semester persistence, and number of enrollments. To analyze these indices, ten classification algorithms were applied to four distinct datasets (BD1, BD2, BD3 and BD4), generated from open data available on this institution data center. Furthermore, optimized versions with Principal Component Analysis (PCA) were used for datasets BD1, BD2 and BD3 for comparison purposes, as they do not incorporate the indices proposed in this study. When comparing the performance of classifiers using the dataset with the indices proposed in this work against the other databases, the Quadratic Discriminant Analysis (QDA) and Naive Bayes models stood out, achieving the best results of accuracy and F1 score, respectively. The critical difference diagram was used to do statistical analyses. The analysis using SHAP (SHapley Additive exPlanations) revealed that the indices persistence and number of enrollments were the most relevant for predicting dropout