Dynamic-GTN: Learning an Node Efficient Embedding in Dynamic Graph with Transformer

Graph Transformer Networks (GTN) use an attention mechanism to learn the node representation in a static graph and achieves state-of-the-art results on several graph learning tasks. However, due to the computation complexity of the attention operation, GTNs are not applicable to dynamic graphs. In this paper, we propose the DynamicGTN model which is designed to learn the node embedding in a continous-time dynamic graph. The Dynamic-GTN extends the attention mechanism in a standard GTN to include temporal information of recent node interactions. Based on temporal patterns interaction between nodes, the Dynamic-GTN employs an node sampling step to reduce the number of attention operations in the dynamic graph. We evaluate our model on three benchmark datasets for learning node embedding in dynamic graphs. The results show that the Dynamic-GTN has better accuracy than the state-of-the-art of Graph Neural Networks on both transductive and inductive graph learning tasks.