Um olhar não supervisionado sobre o desempenho dos participantes maranhenses no Enem 2023

Abstract

Data Science is an interdisciplinary field that can contribute to addressing humanity's challenges in the 21st century, including the search for inclusive, equitable and quality education, particularly in the Brazilian context, where illiteracy still affects millions of people, more severely impacting black individuals in the Northeast Region. The National High School Exam (Enem) has played a significant role in evaluating brazilian basic education and has become the primary gateway to higher education. However, the results of brazilian students, especially in the state of Maranhão, serve as a warning and highlight the need for improvements in basic education. This study is justified by Maranhão's performance in the Ideb and the Enem, placing it among the lowest average results in the country in the 2023 edition. The research problem lies in understanding the variables that play a predominant role in the performance of Maranhão's students in the most important assessment of basic education, from the perspective of educational data mining. Therefore, the general objective is to investigate the factors influencing the performance of Maranhão's students in the 2023 Enem, using exploratory statistical analysis tools and extracting knowledge through unsupervised machine learning, specifically association rule mining with the FP-Growth algorithm. The research approach is quantitative in nature and the methodology employed the CRISP-DM process for knowledge discovery in databases. Computational experiments included the selection of ten attributes through the application of RFE combined with the Random Forest algorithm, as well as the subsequent application of the FP-Growth algorithm in ten trials/configurations on the subset of microdata from Maranhão's participants in the 2023 Enem. The study identified another relevant variable for participants' performance, which enhances the understanding and socioeconomic characterization of Maranhão's population, namely the nature of the occupation of the father or male guardian. Additionally, it revealed underlying patterns in the data not observed in the exploratory analysis, such as the discovery of groups cumulatively affected by different attributes. For example, low performance conditioned by the father's education level and occupation, or the socioeconomic profile characterization of Maranhão's participants based on the selected variables, with a high confidence index. Despite inherent limitations in the quality of the database, the study proved to be socially relevant and has the potential for replication in other regions