Portabilidade de modelos através de transferência de conhecimento para predição do rendimento de alunos de graduação

• Ano de defesa: 2021
• Autor(a) principal: Beltrán, Carlos Antonio Ramírez
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM TECNOLOGIA DA INFORMAÇÃO
• Palavras-chave em Português: Transferência de conhecimento; Aprendizado de máquina; Desempenho acadêmico; Moodle
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/32414

In recent years, one of the major challenges in education has been to predict students’ performance in a certain and reliable way in order to apply different strategies to improve their academic deficiencies. In this way, there are many works focused on finding Machine Learning (ML) models individually, but few use the knowledge acquired from student data of a course degree to predict the students’ outcome of other curses’s degree. Therefore, the main goal of this work is to analise the portability of ML models by using Transfer Learning (TL) in students’ logs extracted form the Moodle system, aiming to predict undergraduate students performance in different courses. Through the applied experimental methodology, each of the two types of groupings formed by the disciplines will be evaluated: those formed according to the undergraduate course and those according to the activities used in Moodle. The data extraction of each group will be performed from the Moodle logs, using the following evaluation methods: cross-validation and hold-out. With this, it will be possible to know whether these evaluations, all performed on the predictive models with the J48 algorithm, tend to show different results in relation to the portability of forecasting models. For evaluation, two scenarios were developed for the execution of experiments, so that each experiment consists of two parts: the choice of models, using the AUC ROC index for Experiment 1 and the F-Measure for Experiment 2; and the validation of the models, using the Precision index for Experiment 1 and the Recall for Experiment 2. The results, even in the evaluation phase, allow us to affirm that it is possible to apply transfer learning between models of the same group in some cases.