The burgeoning field of dynamic graph representation learning, fuelled by the increasing demand for graph data analysis in real-world applications, poses both enticing opportunities and formidable challenges. Despite the promising results achieved by recent research leveraging recurrent neural networks (RNNs) and graph neural networks (GNNs), these approaches often fail to adequately consider the impact of the edge temporal states on the strength of inter-node relationships across different time slices, further overlooking the dynamic changes in node features induced by fluctuations in relationship strength. Furthermore, the extraction of global structural features is hindered by the inherent over-smoothing drawback of GNNs, which in turn limits their overall performance. In this paper, we introduce a novel dynamic graph representation learning framework namely Recurrent Structure-reinforced Graph Transformer (RSGT), which initially models the temporal status of edges explicitly by utilizing different edge types and weights based on the differences between any two consecutive snapshots. In this manner, the varying edge temporal states are mapped as a part of the topological structure of the graph. Subsequently, a structure-reinforced graph transformer is proposed to capture temporal node representations that encoding both the graph topological structure and evolving dynamics,through a recurrent learning paradigm. Our experimental evaluations, conducted on four real-world datasets, underscore the superior performance of the RSGT in the realm of discrete dynamic graph representation learning. The results reveal that RSGT consistently surpasses competing methods in dynamic link prediction tasks.

翻译：动态图表示学习这一新兴领域，因现实应用中对图数据分析日益增长的需求而蓬勃发展，既带来诱人机遇也构成严峻挑战。尽管近期采用循环神经网络和图神经网络的研究取得了令人鼓舞的成果，但这些方法往往未能充分考虑边缘时间状态对不同时间切片间节点关系强度的影响，进而忽视了关系强度波动所导致的节点特征动态变化。此外，图神经网络固有的过平滑缺陷制约了全局结构特征的提取，从而限制了其整体性能。本文提出一种新型动态图表示学习框架——循环结构增强图Transformer（RSGT），该框架首先通过基于任意连续两个快照差异的边缘类型与权重，显式建模边缘的时间状态。由此，将变化的边缘时间状态映射为图拓扑结构的一部分。随后，我们提出结构增强图Transformer，通过循环学习范式捕捉编码图拓扑结构与演化动态的时间节点表示。在四个真实数据集上的实验评估表明，RSGT在离散动态图表示学习领域展现出卓越性能。实验结果显示，RSGT在动态链路预测任务中持续优于现有方法。