Using Machine Learning Models to Evaluate and Predict STEM Postgraduates' Graduation Rates: A Case Study of China

This article aims to use graduation rate data to fit a machine learning model, understand which attributes influence the graduation rate, and predict the graduation rate. The data collection scope includes four independent variables related to graduation rates in 14 subjects of Chinese postgraduates from 2002 to 2022: Total number of doctoral and master's graduates that year, Total number of doctoral and master's Students entrants that year, Total number of doctoral and master's students enrolment that year, Estimated Graduates for Next Year Total. This study assesses five regression models: Linear, Ridge, Support Vector Regression (SVR), Decision Tree, and Gradient Boosting, using metrics such as MSE, RMSE, MAE,
$\mathbf{R}^{2}$
, and K-Fold CV RMSE. Gradient Boosting emerges as the top performer with the lowest MSE (155.976), RMSE (12.489), and MAE (7.494), and the highest
$\mathbf{R}^{2}$
(0.622). It shows robust generalization with a K-Fold CV RMSE of 12.272. In contrast, models like Decision Tree and SVR exhibit significant discrepancies and higher prediction errors. Therefore, Gradient Boosting is the most effective model for predicting graduation rates in subjects like Science, Engineering, Agriculture, and Medicine.