Um processo orientado a dados para geração de modelo de predição de evasão escolar

Abstract

School dropout is an extremely complex problem, as it involves not only a variety of perspectives, but also a variety of different types of dropout behavior. Historically, the most cited school dropout models had their origin in education, however the emerging area of Data Science applied in Education is capable of developing new predictive models, with generally better results when compared to the most used traditional statistical methods. The main objective of this thesis is the proposition of a process for the generation of a predictive school dropout model based on Data Science. To this end, a sequence of steps is defined in order to model an information flow from problem definition to generation of useful information for managers and teachers. The steps consist of: Understanding the Problem, Understanding the Data, Feature Engineering, Feature Selection, Data Balancing, Models, Evaluation and Interpretation. The proposal’s contribution is found in the indication of which techniques and algorithms should be used in each phase of knowledge discovery, and show that the phenomenon of school dropout must be addressed as a problem of imbalanced classes, and should be approached with appropriate tools and metrics, in order to generate a robust and easy to interpret prediction model. The proposed process was validated on educational and socioeconomic data of students Federal Institute of Rio Grande do Norte (IFRN).