The application of machine learning approaches in international large-scale educational assessments: A scoping review

Machine learning approaches have emerged as powerful analytical tools for understanding student learning, which have become increasingly popular in recent years to complement traditional statistical analyses. However, there remains a limited understanding of how machine learning approaches are applied. This scoping review synthesizes 57 empirical studies that applied machine learning techniques to analyze data from international large-scale educational assessments. Specifically, the review explored the types of machine learning approaches used, the research questions addressed, the theoretical frameworks adopted, and the methodological advantages of machine learning approaches. Findings indicated that machine learning approaches, such as random forests, boosting algorithms, and support vector machines, were the most commonly employed approaches. Machine learning was primarily used to select important features, identify key predictors across student subgroups, and benchmark predictive performance across approaches. Notably, 28 out of 57 studies lacked explicit theoretical frameworks, while the remaining studies predominantly relied on macro- and meso-level theoretical frameworks. This review highlights several strengths of machine learning, including its superior accuracy in predicting outcomes, capacity to handle large sets of potentially correlated variables, and ability to generate exploratory insights that complement traditional methods. Concomitantly, we highlight the need for more theory-driven and methodologically robust applications of machine learning approaches in future studies. Given the growing popularity of machine learning, this review has promising implications for researchers who might want to draw on these approaches to advance research on student learning. • Scoping review of 57 studies on the application of machine learning approaches. • Random forest was the most commonly employed approach. • Machine learning was primarily used for feature selection and importance analysis. • Macro- and meso-level theoretical frameworks were mostly used in machine learning-based papers. • Machine learning approaches complemented traditional statistical models.