Modelagem generalista ou individualizada na construção de modelos preditivos para a identificação de insucesso acadêmico

Abstract

The use of technological resources to assist teaching and learning tasks is a reality. The dissemination of virtual learning environments, as a mean of promoting online courses, shows a clear expansion. In addition to tasks that allow the expansion of teaching resources, such systems allow the complete recording of all the interactions of the students inside the courses. This range of information produced can be used to predict at-risk students while the course is taking place, which for educational institutions may represent a reduction in failure and dropout rates. However, the high number of variables involved, especially when several courses are considered, makes it difficult to construct efficient computational models. In this way, this work aims to investigate the construction of generalist models – trained with data from several available courses – counterposing the construction of individualized models – individually trained with data from each course. In this way, a broad set of educational data was extracted, obtained from a higher education institution, composed of different undergraduate programs, courses and academic periods, not using variables that invaded students' privacy. Once the characteristics and transformations of the data that contributed to the identification of academic insuccess during the course were defined, then classical data mining algorithms were applied following both generalist and individualized approaches and to each content unit of the course. The results obtained demonstrate the advantages and disadvantages of both approaches and that given the circumstances the individualized models may be better, obtaining higher hit rates, and that in other circumstances generalist models present a lower cost for the obtaining and maintenance of the predictive models, even with a drop in hit rates.