Sistema preditivo de reprovação de alunos do ensino básico

Abstract

This paper presents the results of developing a student failure prediction system in Basic Education using the evaluation grades from schools affiliated with the Social Service of Industry in Alagoas. Educational performance is a crucial factor in the field of education, making it important to develop a system that can indicate if a student is at risk of failing. The aim of the study was to develop predictive models to estimate the probability of failure in a subject at each evaluation point throughout the year, presenting results with labels indicating whether students are predicted to pass or fail, along with the associated probability. Identifying students at risk can guide pedagogical interventions to improve performance and prevent failures. For the system development, classification algorithms, including K-Nearest Neighbors, Decision Tree, Random Forest, and eXtreme Gradient Boosting, were applied. They were evaluated in terms of performance and functionality concerning the explored solution. Evaluation metrics included accuracy and F1-Score, and the Mann-Whitney test was applied to metric results to determine statistical signi ficance. In this regard, K-Nearest Neighbors showed no significant differences from the Decision Tree but differed from the Random Forest and eXtreme Gradient Boosting. Sta tistically, the Random Forest and eXtreme Gradient Boosting outperformed the others. Additionally, the arithmetic mean was used to synthesize model results, with the Random Forest and eXtreme Gradient Boosting showing the best performance. Consequently, the final version of the predictive system uses the eXtreme Gradient Boosting classification algorithm to generate models and make predictions based on existing data. The choice of eXtreme Gradient Boosting was based on its consistent performance, lower dispersion in relation to obtained averages, and computational efficiency. The results obtained from the chosen approach were satisfactory, with the performance of models generated by eXtreme Gradient Boosting showing approximately 86% accuracy and F1-Score.