Content area
Full text
Predicting those at-risk of dropping out allows schools to assist students before it happens. Machine learning (ML) techniques can predict the likelihood of students completing a course, enrolling in future semesters, or graduating from college. This study compares four ML techniques to predict dropout rates using a student's demographic information and performance in individual courses over all semesters enrolled. Using ten semester models the logistic regression method had the best accuracy of 84.8% versus decision trees (82.2%), neural networks (80.8%), and support vector machines (72.5%). The semester course performance data is a useful input for predicting dropout rates.
Keywords: retention, machine learning, logistic regression, predictive analysis, semester-wide analysis
INTRODUCTION
Retention in higher education is an everlasting battle [NCES, Barbera et al. 2020], just as chum is for companies with subscriptions. Students come from a variety of personal and professional backgrounds. Understanding when a student could be in danger of dropping out is a quintessential problem for colleges. There is not just a monetary incentive, but also a fulfillment of obligation in providing continuing education for those who choose to attend. Identifying students who are in danger of dropping out and taking steps towards their retention could be key in this battle.
In recent years, machine learning methods have become more easily applicable due to wider availability of data and sufficient computing power, which has enabled higher education institutions to develop their own models. Due to the importance of the area, several studies aimed at student retention have been published at dozens of institutions, a lot of them taking advantage of the recent boom in machine learning and relying on its learning methods [Cardona and Cudney, 2020]. What these institutions are predicting with the help of their models can be broadly categorized into three types of studies.
Studies that predict whether a student will:
A.) register for a following semester (i.e. the next semester or the next academic year),
B.) graduate from the institutions sometime in the future,
C.) finish a specific class within a semester.
These types of studies have significant overlap, but each operates with different goals and are not interchangeable. The main goal of predicting whether students register for a close-by semester is to make sure that the registration numbers at a...