Geração de modelos de predição para estudantes em risco de evasão em cursos técnicos a distância utilizando técnicas de mineração de dados

Abstract

Evasion is considered one of the main problems related to e-learning (EAD). In this teaching modality, the interaction between students and teachers is usually mediated by a Virtual Learning Environment (AVA), where the actions taken by students and teachers during the teaching-learning process are recorded in their interaction logs. The large volume of data generated by these interactions allows the use of mining techniques to analyze student data. This work applies data mining and machine learning techniques to logs of students’ interactions of distance technical courses within AVA’s in order to identify students at risk of evasion, using as main input variables for the prediction models Only the count of these interactions and varied attributes of them. Data from logs in the AVA (interaction counts and students’ final situation) of four EAD technical courses were used. Student interactions were counted separately as follows: number of daily interactions, sum of weekly interactions, weekly mean of interactions, standard deviation and final situation. Two different scenarios were evaluated: 1) Generation of prediction models with training and test using data from the course itself and from cross validation and 2) Training of the models with data from three courses and test of the models with data from the course remaining. In the first scenario, prediction models of prediction of ACG evasion up to 84% were obtained before the tenth week of course, reaching 95% until week 52. In the second scenario, most of the prediction models present results of up to 80% In the first ten weeks of the course reaching 98% before the middle of the course. One of the models has achieved an ACG of up to 95% since the first few weeks. In the direct comparison with the statistical model, both techniques showed close results in the first weeks. However, from the tenth week, the models generated through data mining showed a significant growth in the GCA, while the statistical model remained stable. The contribution of this work is the generation of models able to early predict dropout students.