Improving Question Embeddings for Knowledge Tracing with Residual Graph Convolutional Networks

Knowledge Tracing (KT) facilitates the design of personalized learning programs in intelligent education by assessing students' cognitive states and enhancing the efficacy of teaching and learning. The accuracy of KT largely hinges on the quality of the question representations. In recent years, numerous graph-based deep learning approaches, primarily focusing on the interplay between questions and knowledge concepts, have emerged from the KT field. These methods often employ graph convolutional networks to extract question features, yet they commonly overlook the over-smoothing issue inherent in graph neural networks. To overcome this issue, we introduce the Residual Graph Network-based Knowledge Tracing (RGNKT) model, which is designed to preserve critical features in question nodes by incorporating residual connection post-convolution. Additionally, we employ hyper-graph neural networks to enhance the extraction of interactions between questions and knowledge concepts. The experiments of RGNKT on two public datasets demonstrate its capability to accurately predict student performance.