Predictive Insights into U.S. Students’ Mathematics Performance on PISA 2022 Using Machine Learning Models (Poster 49)

• In Event: AERA Poster Session Sunday 1:30 p.m.
  In Poster Session: Machine Learning Approaches to Enhancing Learning Environments With Artificial Intelligence

Abstract

This study aims to identify the influential predictors of math performance using machine learning (ML) models and compare them to those identified by traditional literature review methods. Using the PISA 2022 U.S. dataset, we employed three ensemble tree-based ML models (Random Forest, XGBoost, and LightGBM) on 143 derived variables. XGBoost demonstrated the best prediction accuracy (rMSE = 69.82) and efficiency (training time = 4.14 seconds), identifying 10 significant predictors, including math self-efficacy, ESCS, and familiarity with math concepts. Accumulated local effects plots visualized these predictors’ impact on math performance. Our analysis showed that predictor selection by machine learning was superior to that from literature review, as confirmed by a Wilcoxon signed-rank test. The implications of these findings are discussed.

Authors

• Li Zhu, University of Iowa

• Hyesun You, University of Iowa

• Minju Hong, Chung-Ang University

• Zhenhan Fang, University of Iowa