HRMNN: Heterogeneous Relationship Mined Graph Neural Network

Qianyi Zhan, Jing Wang, Ying Li, Zhenping Xie, Yuan Liu

Abstract: Heterogeneous graph embedding has attracted considerable attention in the field of graph data analytics. It can effectively capture relations among different types of nodes. However, existing methods have confronted with several challenges. On the one hand, a number of approaches rely on prior knowledge of meta-paths. On the other hand, some of them ignore the influence of dataset preprocessing on the graphs. Moreover, certain algorithms tend to omit the information of intermediate nodes along the meta-path when aggregating node relationships. To solve these problems, we propose a heterogeneous graph embedding model named Heterogeneous Relationship Mined Graph Neural Network (HRMNN). The model incorporates a relational graph generator that effectively utilizes the topological attribute of heterogeneous graph, and combines object-level aggregation and multi-head attention mechanism to generate richer node representation. Experiments such as node classification and link prediction on multiple datasets demonstrate that HRMNN has better performance compared to state-of-the-art baselines.