Predicting Student Attrition in University Courses

Educational institutions are actively engaged in extensive initiatives to mitigate student dropout rates. In addition to various strategies, the integration of machine learning (ML) models stands out as a prominent tool. These models play a crucial role in pinpointing students who are at risk of discontinuing their studies and offer tailored interventions to assist them in catching up with their academic coursework. Developing effective educational machine learning models entails consideration of specific features. Typically, the aim is to leverage a large sample size and incorporate an extensive set of predictors for optimal model construction. While the criterion for a substantial sample size is often met in higher education or degree-level analyses, identifying, and incorporating effective predictors pose challenges. In the realm of education, the prospect of building a reliable model is enhanced when predictors are tailored to subjects or even specific curricula. This customization is justified by the inherent differences in content, student enrollment, and requirements across various university courses. However, a critical challenge emerges as the use of course-specific predictors proliferates—it leads to a subdivision of existing observations into numerous parts. Consequently, this segmentation diminishes the size of the sample available for model training, posing a potential limitation to the robustness of the machine learning model. Predicting within a semester also poses special problems, since the observations that follow each other in time are not independent of each other. This chapter is about building ML models of university courses to better handle the prediction of student attrition.