End-to-End Deep Knowledge Tracing by Learning Binary Question-Embedding

Abstract

Recent advancements in computer-assisted learning systems have increased research in the are of knowledge tracing, which estimates student proficiency based on their past interaction with the learning systems. In this context, the method called Deep Knowledge Tracing (DKT), which leverages recurrent neural networks, shows remarkable performance; however, existing knowledge tracing methods, including DKT, require human-predefined skill-tags that show what skill is required to solve a question. This limits the optimization of modeling student proficiency and the application to real-world data, which are often not well-organized by skill-tags. In this paper, we propose the first end-to-end DKT model (E2E-DKT), which does not depend on any human-predefined skill-tags. Regarding the correspondence between questions and tags as a binary embedding matrix, we introduce a new Q-Embedding Model, which learns the embedding matrix to help predict student proficiency from the student question-answer logs only. We provided two techniques for learning a better matrix, one is the reconstruction regularization of question-space and tag-space, and the other is the sparse regularization of the question-embedding matrix. Using two open datasets, we empirically validated that the learned tags show the same or better performance on DKT and have an information-efficient structure and more hierarchical relations among each tag than the human-predefined tags. These results show the potential of our proposed method to enhance the applicable scope and effectiveness of DKT, which could help improve the learning experience of students in more diverse environments.