Analyzing feature importance for a predictive undergraduate student dropout model

Alberto Jiménez-Macias ; Pedro Manuel Moreno-Marcos ; Pedro J. Muñoz-Merino ; Margarita Ortiz-Rojas ; Carlos Delgado Kloos

Computer Science and Information Systems, Tome 20 (2023) no. 1

Résumé

Worldwide, one of the main concerns of universities is to reduce the dropout rate. Several initiatives have been taken to avoid this problem; however, it is essential to recognize at-risk students as early as possible. This article is an extension of a previous study that proposed a predictive model to identify students at risk of dropout from the beginning of their university degree. The new contribution is the analysis of the feature importance for dropout segmented by faculty, degree program, and semester in the different predictive models. In addition, we propose a dropout model based on faculty characteristics to try to infer the dropout based on faculty features. We used data of 30,576 students enrolled in a Higher Education Institution ranging from years 2000 to 2020. The findings indicate that the variables related to Grade Point Average(GPA), socioeconomic factor, and a pass rate of courses taken have a more significant impact on the model, regardless of the semester, faculty, or program. Additionally, we found a significant difference in the predictive power between Science, Technology, Engineering, and Mathematics (STEM) and humanistic programs.

Détail

Keywords: dropout model, features importance, data mining, learning analytics