Large-scale educational assessment studies (LSAs) regularly collect massive amounts of very rich cognitive and contextual data of whole student populations. Currently, LSAs are limited to reporting student proficiencies in the form of plausible values (PVs). PVs are random draws from the posterior distribution of a student’s ability, which is based on the Bayesian approach with the prior distribution modeling the student background within the population and the likelihood test item response using the Rasch model. While PVs have shown to be a reliable estimate for proficiencies of populations, a more comprehensive study of these rich data sets by deploying machine learning algorithms may provide a better understanding of the underlying factors affecting student performance and thus yield to better and more interpretable predictive models. This paper presents such a novel approach to learn directly from LSA data by deploying a combination of both unsupervised and supervised learning feature selection algorithms to predict student performance on math scores. Our technique learns the difficulty level of different math questions and predicts weather or not a student with a particular background profile will be successful in answering correctly.