Predicting Dropout Using High School and First-semester Academic Achievement Measures

Due to the big data accumulated in educational administrative systems and due to the advance of machine learning techniques, a new scientific discipline has emerged in the last few years, namely educational data science. An important research objective of this field is to predict dropout and improve graduation rates, in particular in STEM higher education.The goal of this study is to identify students at risk of dropping out at a large Hungarian technical university using predictive analytical tools. We use data of 10,196 students who finished their undergraduate studies (either by graduation or dropping out) between 2013 and 2018. We analyze dropout predictability in two main scenarios: first using data available at the time of enrollment, e.g. pre-enrollment achievement measures, and personal details, then in another scenario we supplement this feature set with first-semester performance indicators and use this richer set of attributes for further analysis. We apply artificial neural networks and boosting algorithms for prediction, and examine how the predictive power can be improved by the additional information. In other words, we study the incremental predictive validity of the early university performance indicators on graduation over the pre-enrollment achievement measures and vice versa.