Graph neural network-based knowledge tracing model integrating student-exercise interaction and knowledge-exercise relationships

Abstract

The growing importance of knowledge tracing (KT) in online education has led to the development of various KT methods. Recent studies have focused on exploiting the relationship between knowledge concepts (KCs) and exercises. However, existing methods are limited in integrating student-exercise interactions and exercise-KC relationships, and inferring on KCs and exercises that were not seen dur- ing training. To address these challenges, we propose Dual attention Graph-based Knowledge Tracing (DGKT), a GNN-based model that integrates exercise-KC hypergraphs and student-exercise bipartite graphs, without relying on sequence models such as RNN. Moreover, our model can adapt to evolving graphs by learning through student-exercise subgraph units. It incorporates multiple KCs through the use of dual attention graph neural networks, allowing for the importance of connected neighboring nodes and the importance of nodes within a subgraph to be identified. Additionally, DGKT is able to handle multiple KC types simultaneously and improve explainability by utilizing the importance of each node through a global attention layer. Our proposed model outperforms current state-of-the-art models, as demonstrated by experimental results on multiple datasets.