Explorando curvas principais com a meta-heurística lobo cinzento para a classificação de dados sintéticos e de desempenho acadêmico dos estudantes no ENEM

Abstract

Education is fundamental to a country’s development, and in Brazil, the need for improvements that can be driven by the use of Information Technology stands out. Through the National High School Exam (ENEM), the country’s largest educational exam and one of the main gateways to higher education, it is possible to assess aspects of the quality of education by constructing educational indicators based on student performance on the exam. To predict student performance on the exam based on variables present in the ENEM database, such as socioeconomic information, school characteristics, and participation data, Machine Learning (ML) techniques have been increasingly used in this context, allowing the identification of performance patterns, possible irregularities, and the customization of pedagogical strategies. The problem of predicting student performance on the ENEM has been investigated by several authors, but many have not explored other ML techniques, such as Principal Curves (PC). In recent years, the PC method has been applied in various areas, demonstrating potential in classification problems. In this context, this research aims to apply the K-segment PC extraction method to classify the academic performance of students who took the 2023 ENEM exam, considering class 0 for students who did not present the expected performance and class 1 for those with good performance on the exam, and to evaluate it on synthetic bases. Furthermore, the GreyWolf Optimizer (GWO) hyperparameter optimization method was applied to automatically determine the hyperparameter values for the K-segment PC method, a task that can be determined manually but is complex. The methodology comprises the steps of preparation, dimensionality reduction, class balancing, and transformation of the input variables from the 2023 ENEM database, including socioeconomic variables, school characteristics, and participant information, in addition to applying the GWO technique to optimize the model hyperparameters. The classification methods were evaluated using metrics such as accuracy, F1-Score, precision, recall, and Kappa coefficient. In experiments conducted on synthetic datasets, the method showed good performance in compact, elongated, spherical, and spiral datasets, with metrics greater than 0.9700. In experiments conducted using the ENEM 2023 database, the PC approach showed competitive results compared to the literature-referenced methods (Extreme Learning Machine, Naive Bayes, and Random Forest). Among the evaluated approaches, the worst result was observed with t-SNE, with an accuracy and recall of 0.7310, precision of 0.7314, F1-Score of 0.7309, and a Kappa coefficient of 0.4621 in the test set. While the Select K-Best method obtained the best results, with an accuracy and recall of 0.7603, precision of 0.7612, F1-Score of 0.7601, and a Kappa coefficient of 0.5206 in the test set, it outperformed the Naive Bayes method in this configuration. These results indicate that the proposed approach is promising for classifying academic performance, especially when combined with appropriate dimensionality reduction and optimization techniques. Keywords: Educational Data; Academic Performance; ENEM; Machine Learning; Pattern Recognition; Principal Curves; K-segments.