Predicting Student Performance: The Case of Combining Knowledge Tracing and Collaborative Filtering

In the past few years, many competing learning models have been proposed for improving the accuracy of predicting student performance (PSP). A well-studied subclass of algorithms focused on PSP uses temporal models to determine the knowledge state of users. Bayesian Knowledge Tracing (BKT), as one of the leading models in this subclass, uses Hidden Markov Models to capture the student knowledge states. An emerging new subclass of algorithms focused on PSP uses collaborative filtering, which is used primarily by recommender systems. Matrix Factorization (MF), a leading model in this subclass, can be presented as a rating prediction problem where students, tasks, and performance information are mapped to users, items and ratings, respectively. BKT and MF complement each other’s strengths and limitations quite effectively. In particular, BKT relies on four skill-specific parameters for learning the sequential behavior of learners on each concept, but it does not capture the similarities among users and items. In contrast, MF uses latent factors to exploit the similarities among users and items from learner-item performance, but disregards any temporal effect in modeling student learning. In this paper, we aim to investigate the effect of combining variations of BKT and MF using a proposed algorithm that exploits the power of MF in modeling the implicit similarities among learners and items while using the explicit parametrization of BKT towards improving PSP. Our results on four benchmark educational datasets show that our approach outperforms the base classes as well as traditional techniques such as linear regression, logistic regression and Neural Networks for combining BKT and MF