A Knowledge Tracing Model Based on Graph Attention Mechanism and Incorporating External Features

In recent years, research has focused heavily on Knowledge Tracing (KT), a crucial technique for learner state modeling in intelligent education. There are several KT models based on graph convolutional networks (GCN-KTs), but none of them can distinguish the importance of exercises or knowledge concepts. Existing GCN-KTs treat all neighboring nodes “equally” when performing graph convolution operations for exercise and concept embeddings, resulting in insufficient robustness of the generated node representations. Based on GCN-KTs, we offer a Knowledge Tracing model that is based on the Graph Attention Mechanism (GAFKT) and has an encoder-decoder structure. The encoder applies the self-attention layer to the topology and node features of the exercises and concepts, respectively, and the decoder uses the inner product to reconstruct the graph structure. In GAFKT, the semantic model of student knowledge and exercises is further enriched by mapping the external feature embeddings in the original data to the embeddings of the exercises and concepts in the same space. This work was experimented with several the most advanced models on two open source datasets for comparison, and the results demonstrated the effectiveness of GAFKT.