Uma abordagem de game learning analytics para identificação de perfis comportamentais na utilização de jogos educacionais

Abstract

Currently, research that seeks to evaluate the learning acquired by players from a Serious Game has been adopting measures to demonstrate evidence collected in realtime, using techniques such as those in the areas of Machine Learning and Deep Learning. However, few studies seek to carry out this type of evaluation and techniques in Serious Games for early childhood education. That said, this study sought to apply a Game Learning Analytics approach that has two complementary objectives: 1) Identify behavioral patterns; 2) Predict the acquired learning effect. For this purpose, this research employed data collected by digital games to assess how different students benefited from the Escribo Play word reading and writing intervention during a trial with 749 preschool students. For behavioral profiles identification, a cluster analysis was performed to form groups, the KruskalWallis method to understand if there are differences between the groups and the Effect Size, to reveal how different they are. For the prediction of the learning effect, 4 classification algorithms were trained and validated from the set of combinations of interaction variables collected in the games. From the cluster analysis process, three behavioral profiles were identified that present different performances: Profile 1, with the largest number of students, presented the expected performance for this intervention; profile 2 presented the largest effectsize, being a reference for engagement with this intervention; and profile 3, which was formed by two groups, which due to the small effectsize, seem to represent children who were not yet ready to benefit from the educational intervention. In the process of training and validating the classification algorithms, we obtained as best results an Accuracy of 74% and Precision of 81%, in the classification of student performance, results that are within the expected for the context of early childhood education. As a result obtained in these experiments, we found that the best way to differentiate students from each other is through the interaction characteristics that represent the errors made during the use of these games. Even though there is a notable lack of studies that address early childhood education, the results shown here are promising and indicate that it is possible to research and further explore the use of Game Learning Analytics techniques in this context.