Learning analytics: Collaborative filtering or regression with experts?

An intelligent learning analytics based on an enormous amount of education data is a key enabler of the next generation of education; among many tasks of the intelligent learning analytics, personalized prediction of test responses based on the record of each individual learner is of the utmost importance. In recent years, a variety of machine learning algorithms for predicting test outcomes have been proposed, and two of the most prominent approaches are collaborative filtering and logistic regression. Collaborative filtering is fully data-driven since it does not require any extra information other than test outcomes while logistic regression is applicable only when questions can be independently analyzed by experts. In this work, we first propose a new model for test responses, and propose a collaborative filtering algorithm with enhanced human-interpretability based on the new model. Then, we evaluate the prediction performance of these approaches using a large education data set, collected via mobile applications for English Language Learning. Our experimental results show that the fully data-driven collaborative filtering approach can predict test outcomes better than the logistic regression approach.