Aprendizagem de máquina para predição do abandono e evasão dos estudantes do Estado do Ceará.

• Ano de defesa: 2023
• Autor(a) principal: Teles, Igor Antônio Gomes
• Orientador(a): Fernandes, Carlos Alexandre Rolim
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Evasão Escolar; Violência; Machine Learning; Predição
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/73132

Leaving and dropping out of school are frequent themes in Education. The numbers give an idea of the size of the problem. In 2018, around four out of ten 19-year-old Brazilians did not finish high school based on the Continuous National Household Sample Survey (PnadC), by IBGE. Dropout occurs when the student stops attending classes during the school year. School dropout, on the other hand, concerns the situation of the student who dropped out of school or failed in a given school year, and who in the following year did not enroll to continue his/her studies. The purpose of this project is to propose models for predicting dropout and dropout situations for students in the state of Ceará, using social databases, school performance and mothers' records in the CVLI and Maria da Penha databases. Another purpose of the work is to determine which factors have the most impact on evasion and abandonment. Longitudinal data from the years 2012 to 2019 of school data obtained from the School Census were used to verify the situation of students who dropped out or dropped out. In total, 4 databases were used: School Census, SPAECE, CVLI and Maria da Penha. The procedures were carried out through the Postgresql database management system, SPSS Software and Weka. After pre- processing, cleaning and applying filters, the data were used for machine training and prediction verification for decision-making about possible situations of evasion and abandonment. The Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Random Forest classifiers were used, Correlation based feature selection - CFS was also applied to find the best attributes for the study, with performance in Portuguese and Mathematics, ethnicity being selected as attributes. , teaching stage and the indicator of the mother's presence in bases of violence. The respective accuracies of 83.9%, 78.24% and 71.4% were achieved, which concludes that the MLP classifier obtained the best result.