Topological Network Field Preservation for Heterogeneous Graph Embedding

Heterogeneous graph (HG) embedding, aiming to represent the nodes in the graph as a low-dimensional vector form for further reasoning to better implement downstream tasks, has attracted considerable attention in recent years. Most existing HG embedding methods use the meta-paths to preserve the proximity or adapt graph neural networks (GNNs) to facilitate the message-passing process. However, these methods neglect to analyze the shape properties of nodes and the influence of each node from a topological perspective, thus cannot fully explore the information on higher-order connectivity of HG and be effectively support more complex tasks of network analysis. In this paper, a novel HG embedding model (TNFE) is proposed to capture the topological link structure and the higher-order interactive information between nodes simultaneously. Specifically, persistent homology is used to reconstruct the connection between nodes in HG. Then the neighborhoods of the nodes are aggregated based on a graph convolutional network. Moreover, modular topology centrality is defined to sample the topological network field structure of each node. Finally, multi-task learning task is built to preserve the topology connectivity and the topological network field proximity simultaneously. The extensive experiments on three real-world datasets show that our method outperforms the state-of-the-art approaches on node classification and clustering task.