Graph learning considering dynamic structure and random structure

Graph data is an important data type for representing the relationships between individuals, and many research works are conducted based on graph data. In the real-world, graph data usually contain rich time information, giving rise to the concept of dynamic graphs, which contain the temporal evolution information. However, current dynamic graph methods tend to focus on capturing low-order neighborhood information (first-or second-order) and rarely consider higher-order global information, due to the constraints imposed by dynamic graph loading. This can limit their ability to gain a good global sense of the field. To address this challenge, we propose a novel method considers Dynamic Structure and Random Structure, called DSRS, which introduces the classic random walk technique to mine high-order structural information. Specifically, we introduce the node2vec algorithm as a pre-train model to perform random walks on the graph for initial embeddings generation, and then utilize contrastive loss to align the temporal node embeddings focusing on dynamic structure with the random node embeddings focusing on high-order structure, constraining the model to maintain balance between the two different structural information. We performed experiments on link prediction using various real-world dynamic graphs, and the results showed a maximum improvement of 4.34% compared to existing temporal graph learning methods. Furthermore, we conducted experiments to test the sensitivity of our parameters, as well as ablation experiments, which both confirmed the effectiveness of our proposed feature pre-training and parameter fine-tuning methods. (c) 2023 The Authors. Published by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).