Accuracy-aware Deep Knowledge Tracing with Knowledge State Vector Loss

Abstract

In major e-learning platforms such as intelligent tutoring systems (ITSs) and massive open online courses (MOOCs), the students are often recommended what course materials to take based on their past interactions. Knowledge Tracing (KT) is the task of modeling students' academic abilities. Given a sequence of student's learning history, it predicts how well they will perform in the next interaction. Deep Knowledge Tracing (DKT) uses a recurrent neural network (RNN) to capture the underlying structure of the student's understanding. In this paper, we point out the ​accuracy rate problem that the model won't reproduce the accuracy ratio. This is a limitation of the existing loss function in DKT that it only learns the probability of correctly answering a problem in the next interaction. We introduced the ​Knowledge State Vector loss​, which captures the accuracy rate of all knowledge concepts, to measure and train the model.