Leveraging machine learning techniques to analyze computing persistence in undergraduate programs

Abstract

Although student retention remains a significant concern for all Science, Technology, Engineering, and Mathematics (STEM) fields, it is particularly problematic in computing, where enrollment in such programs has not kept pace with the industry demands. Thus, finding meaningful patterns in historical data can help education researchers to reveal the possible reasons for students' withdrawal from a university, and can provide guidelines and mechanisms that lead to improving retention rates. To achieve this goal, we considered the importance of different factors in the graduation of computing students, and generated a predictive model for student graduation using the Multiple-Institution Database for Investigating Engineering Longitudinal Development (MIDFIELD) dataset. We observed that considering input and environment educational variables, cumulative GPA, number of terms registered, start year, institution, and being a transfer student, are the most important features, respectively. Our results also demonstrate that Random Forest algorithm produced a more accurate result on this dataset compared to other machine learning algorithms. We anticipate findings from this ongoing work will give insight to the computing education community and researchers to better understand the relative success of computing students, and that this will, in turn, enable strategic solutions to attain higher retention rates.