Interpretable Machine Learning Models for PISA Results in Mathematics

The Program for International Student Assessment (PISA) 2022 provides a global framework for assessing educational performance worldwide. We addressed a principled observation of characteristics and disparities in Mathematics performance among Spanish adolescents while showing the factors influencing these results. To address this question, we proposed using advanced Machine Learning techniques through possibly non-linear predictive models that identify key drivers of Mathematics performance to inform data-driven educational policies and interventions that improve learning outcomes. By preprocessing the PISA dataset, we categorized students into Low, Medium, and High proficiency levels and employed various binary classification models to discern predictive patterns. In addition, a stacking meta-model integrating the strengths of eight distinct predictive models was developed to enhance prediction accuracy. Our results demonstrated that the meta-model outperforms individual models in predicting student performance across various proficiency levels, consistently showing superior metrics in Precision, Recall, and Area Under Curve (AUC) scores. Specifically, the meta-model achieved an AUC score of 0.9766 when classifying students in the Low and High proficiency categories. We adopted the Shapley Additive exPlanations method to demystify the model decisions, highlighting significant predictors such as grade repetition, access to digital devices, and extracurricular Mathematics classes. We also introduced an interactive dashboard, harnessing Uniform Manifold Approximation and Projection for dimensionality reduction and enabling a granular view of the educational landscape. The intention of all this is to contribute to an education system that is well-informed and effectively adapts to the diverse needs of students.