Big Data Analytics para a classificação do risco de abandono escolar em cursos do ensino superior

Abstract

School dropout is a constant concern of higher education institutions, with divergent and complex factors reported by various authors. We can note that this problem is quite broad and exploited, since it is possible to find specialized sectors or assistance programs such as psychologists, housing assistance, monitoring programs, among others, in order to minimize the number of students dropping out. However, these proposals depend on the student himself to seek necessary help, opening a gap for those who do not feel comfortable to seek or do not have full knowledge of the case itself. This work proposes a model for the prior identification of dropouts, with the aim of making higher education institutions able to understand cases of dropout and, if possible, refer them to specialized sectors. For that, we use as a case study the Instituto Politécnico de Bragança, which provided 200 million records related to students enrolled between 2008 and 2017. We analyze and process the Big Data provided for molding it as input parameters of machine learning algorithms. We first tested three algorithms and found that random forest proves to be the most efficient in this context. From this, we take advantage of the data volume to identify which would be the best training cycle and obtain that the period of 4 years can achieve better results. In the improvement of the model we added two more attributes to highlight the trajectory of the student. For the implementation and visualization of the model, we developed a data extraction tool and a Web application, which through different levels of access, besides being able to identify students at risk of abandonment, also allows users to make comparative analysis between schools and courses by of a custom page with statistics transformed into graphs and tables. The study presents as a good solution for the prior identification of students at risk of dropout, enabling analysis and referrals. The model can still be extended to more parameters and tends to obtain better results over the years by improving by enhancing the attributes created.