Uso de um método preditivo para inferir a zona de aprendizagem de alunos de programação em um ambiente de correção automática de código

Abstract

CS1 (first year programming) classes are known to have a high dropout and non-pass rate. Thus, there have been many studies attempting to predict and alleviate CS1 student performance. Knowing about student performance in advance can be useful for many reasons. For example, teachers can apply specific actions to help learners who are struggling, as well as provide more challenging activities to high-achievers. Initial studies used static factors, such as: high school grades, age, gender. However, student behavior is dynamic and, as such, a data-driven approach has been gaining more attention, since many universities are using web-based environments to support CS1 classes. Thereby, many researchers have started extracting student behavior by cleaning data collected from these environments and using them as features in machine learning (ML) models. Recently, the research community has proposed many predictive methods available, even though many of these studies would need to be replicated, to check if they are context-sensitive. Thus, we have collected a set of successful features correlated with the student grade used in related studies, compiling the best ML attributes, as well as adding new features, and applying them on a database representing 486 CS1 students. The set of features was used in ML pipelines which were optimized with two approaches: hyperparameter-tuning with random search and genetic programming. As a result, we achieved an accuracy of 74.44%, using data from the first two weeks to predict student final grade, which outperforms a state-of-the-art research applied to the same dataset. It is also worth noting that from the eighth week of class, the method achieved accuracy between 85% and 90.62%.