Projecting Incoming Cohort Size and Course Enrollments via Machine Learning

Session Abstract: The Admitted Matriculation Projection (AMP) is a supervised machine learning model (LightGBM) that predicts the size, characteristics, and course enrollments of the incoming Fall cohort up to 8 months prior. Using weekly snapshots of historical application activity, AMP compares each student admitted for the Fall term to contemporaneous data on prior cohorts to generate individualized probabilities to enroll in ~75 lower-level courses. AMP’s granularity (student-course pairs) yields greater accuracy and flexibility than standard methods extrapolating historical aggregate data. AMP also accounts for fluctuations in application activity and projects cohort characteristics, such as demographic composition, academic background, and geographic distribution. We highlight AMP’s use of advanced, free, open-source Python packages like FLAML (automated machine learning), MiceForest (missing data imputation), and SHAP (model explainability) that could assist a wide range of predictive IR tasks.

Keywords: machine learning, AI/ML, decision support, resource allocation and planning