On the use of Bayesian probabilistic matrix factorization for predicting student performance in online learning environments

Thanks to the advances in digital educational technology, online learning (or e-learning) environments such as Massive Open Online Course (MOOC) have been rapidly growing. In the online educational systems, however, there are two inherent challenges in predicting the performance of students and providing personalized supports to them: sparse data and the cold-start problem. To overcome such challenges, this article aims to employ a pertinent machine learning algorithm, the Bayesian Probabilistic Matrix Factorization (BPMF) that can enhance the prediction by incorporating background information on the side of students and/or items. An experimental study with two prediction scenarios and 24 experimental conditions was conducted to study the BPMF based on real online learning data. The results show that the lower rate of missingness and the appropriate dimensionality of latent features provided better prediction accuracy in both prediction scenarios. The use of side information enhanced the prediction accuracy but the effect was diminished for the high dimensional latent features when the data are sparse. The methodological value, applicability, and practical implications of the BPMF and side information to the online educational systems were also discussed.