Graph Attention-Enhanced Knowledge Tracing: Unveiling Exercise Variability and Long-Term Dependencies

Knowledge tracing (KT) is vital for predicting students' mastery of knowledge concepts (KCs) based on interactive data from their practice sessions, facilitating adaptive learning resource recommendations. Despite the effectiveness of existing models, challenges persist, such as handling long-term dependencies in RNN-based models and real-time tracing of knowledge states across various KCs in attention mechanism-based models. Additionally, latent information within exercises and variations in difficulty levels among exercises under the same KC are often overlooked. To address these challenges, we propose a Graph Attention Mechanism-based Knowledge Tracing model (GAKT). Our approach involves constructing a bipartite graph to represent the exercise-KC relationship, integrating a difficulty vector into the interaction vectors of exercises, and utilizing Graph Attention Networks (GAT) to extract embeddings. Incorporating an attention mechanism, our model features a knowledge state revival module that captures long-term dependencies and provides real-time tracing of students' knowledge states. Extensive experiments demonstrate that GAKT outperforms the state-of-the-art graph-based KT model (GIKT) by an average of 2.3% in terms of AUC across three datasets, showing the importance of high-order information for deeper associations between KCs and exercises in the KT task. This study contributes to advancing knowledge tracing methodologies in the field.