Estimation of Persistence at a Community College: A Comparison of Alternative Machine Learning Models

Abstract

This research focuses on developing persistence models for Rio Salado College. It is an initial effort to predict persistence from one term to the next. Several ensemble models are experimented and compared in their respective key metrics such as: confusion matrix, AUC, F1-Score, and feature importance. Exploratory data analysis is undertaken to narrow the set of variables utilized in the models. Two models were considered for possible implementation: a logistic regression and a gradient boosting machine. The former is easier to implement and explain to non-technical personnel, while the latter behaves like a black box. Based on key performance metrics, the model of choice was the gradient boosting machine. Development and testing were conducted with python using jupyter notebooks. The author hopes that this experimental process will fill a vacuum in the analytical needs of community colleges.