OSGNN: Original graph and Subgraph aggregated Graph Neural Network

Abstract

Heterogeneous Graph Embedding (HGE) is receiving a great attention from researchers, as it can be widely and effectively used to solve problems from various real-world applications. The existing HGE models mainly learn node representation directly on the whole heterogeneous graph by aggregating neighboring information, which unavoidably leads to the loss of useful high-order information. Another mainstream is to split heterogeneous graphs into different homogeneous subgraphs and then learn representations separately. However, this isolated handling way is prone to the loss of important interactions between the nodes of the same type. To address the above challenging but interesting problems, we propose an Original graph and Subgraph aggregated Graph Neural Network (OSGNN). Specifically, we first split the original heterogeneous graph into several subgraphs, and then weighted combine them to get a new meaningful homogeneous graph. Finally, the first-order and high-order information of the target node are learned from the original heterogeneous graph and the homogeneous subgraph respectively and concatenated as the final node representation. Extensive experiments on three real-world heterogeneous graphs demonstrate that the proposed framework significantly outperforms the state-of-the-art methods. The source codes of this work are available on https://github.com/ZZY-GraphMiningLab/OSGNN.