Math proficiency prediction in computer-based international large-scale assessments using a multi-class machine learning model

This study examines whether a multi-class predictive machine learning model can be used to predict students’ math performance in computer-based international large-scale assessments. Raw data from the PISA 2012 computer-based assessment was used to conduct the experiments. Three-class neural network and random forest machine learning models were developed and compared in prediction performance for students’ proficiency in mathematics. The three levels of math proficiency were low, mediocre and high. Both machine learning models relied on background variables from student, parental and school questionnaires, as well as on the student math assessment data. Manual and a few automatic feature selection methods were explored while developing the machine learning models. Due to the imbalance in the three-class dataset Cohen’s Kappa coefficient and ROC AUC were used as main metrics of the model performance.