Recurrent Neural Collaborative Filtering for Knowledge Tracing

In knowledge tracing (KT), a computer system infers the skill level of a student from their interaction with coursework. This work introduces a new method to KT called recurrent neural collaborative filtering (RNCF) that can separate student learning and task difficulty traits into distinct parameter sets. Using five KT data-sets, with binary and scalar response modes, we show this method can improve upon previous predictive approaches. We illustrate the method’s ability to cluster students and exercises into like groups, a result that bears promise for research into personalised learning and curriculum improvement.