Uma Abordagem Baseado em Ranking para a Predição de Evasão Escolar: Um Estudo de Caso na Universidade Federal de Mato Grosso do Sul

Abstract

Academic dropout is a problem discussed in several educational institutions, as it entails an expense for the institution and a significant educational loss for the student. In addition, causes defined by human perceptions can be incorrect and consequently lead to wrong preventive actions. On the other hand, the successful use of data mining techniques is observed in several educational institutions in Brazil and in the world for the automatic extraction of hypotheses and the generation of models to classify whether a student has a propensity to drop out or not. The use of classification in this scenario can generate unsatisfactory results, especially when the classification confidence between the evasion and non-evasion classes are close. Another point to consider is when there is an imbalance between classes, known to generate inaccurate classification models in the minority class. In addition, the causes of dropout may vary by location and course. Alternatively, ranking test instances based on their propensity to belong to a certain meaningful class and ranking them would more precisely direct effort to students most likely to drop out. Given that, this work presents a method for detecting dropout based on ranking and having as a training set the data from students’ school records. It is then evaluated whether dropout occurs in the top ranked students with dropout potential. Furthermore, the method was used in different courses and different areas of knowledge at the Federal University of Mato Grosso do Sul (UFMS) to evaluate its behavior in different scenarios. The results show an accuracy of 0.94, 0.98 and 0.96 for the first 50 ranked students in the exact, biological and humanities areas. And for the first 10 the precision was from 0.99 to 1.0. On the other hand, with the use of classification models, the results obtained were 0.72, 0.79, 0.78 for accuracy and 0.68, 0.65, 0.64 for precision. It is worth mentioning that despite the ranking evaluation only encompassing the top-k students in the dropout class, the assertiveness in the classification makes the efforts better targeted and hopefully more effective. Therefore, given the assertive prediction of the proposed method for students with greater dropout potential, those interested will be able to apply more efficient strategies to combat dropout.