Multi-Scale Variational Graph Auto-Encoder Based on Subgraph Learning

Link prediction is a crucial component in the field of graph data analysis, playing a vital role in various disciplines such as social network analysis, construction of biological information networks, and research on information dissemination networks. This paper proposes a multi-scale variational graph auto-encoder model combined with subgraph partitioning, SGMS-VGAE, which aims to capture the structural features of graph data by integrating subgraph learning and a multi-scale variational graph auto-encoder framework. The model effectively extracts structural features of graph data through static subgraph partitioning and multi-scale variational inference techniques. Moreover, by employing subgraph partitioning techniques, the model can filter out noise and redundant information, focusing on local structures, thereby accelerating the training process. To validate the effectiveness of the SGMS-VGAE model, extensive experiments were conducted on real-world datasets such as Cora, Citeseer, and PubMed. The experimental results demonstrate that SGMS-VGAE outperforms existing methods in link prediction tasks.