Multiscale Knowledge Tracing With Concept Map Representation

Knowledge Tracing plays an important role in intelligent educational systems by assessing the state of students' knowledge to build models that predict their future performance. Existing KT models have made significant progress in predicting students' future performance, but they ignore the influence of question difficulty on the state of knowledge, especially at different fine-grained levels (e.g., overall knowledge state and concept-level knowledge state). In addition, existing models less consider the similarity between concepts and the transfer of prior knowledge, which limits their ability to accurately capture the change of knowledge states. For this reason, we propose a new model called Multiscale concept graph knowledge tracing (MCGKT) that utilizes educational theory to model student processes. On the one hand, an overall student knowledge tracing and a more detailed concept state were designed to explore the impact of question difficulty level on knowledge states at different scales of granularity. On the other hand, a Concept Map Neural Network (CMNN) was designed to utilize the prior, similarity relationships between concepts to convey the interactions between concepts. In this paper, extensive experiments on four real data sets demonstrate that the predictive accuracy of the Multiscale Concept Map Knowledge Trace model is superior to existing models. In addition, the MCGKT provides an interpretable framework for investigating the influence of question difficulty on knowledge states.