HMSG: Heterogeneous graph neural network based on Metapath SubGraph learning

Abstract

Heterogeneous graph neural network (HGNN) models, capable of learning low-dimensional dense vectors from heterogeneous graphs for downstream graph-mining tasks, have attracted increasing attention in recent years. For these models, metapath-based methods have been widely adopted. However, most existing metapath-based HGNN models either discard intermediate nodes within a metapath, resulting in information loss, or indiscriminately aggregate information along a metapath containing different types of nodes, resulting in unavoidable learning bias. To overcome these limitations, a new HGNN model named HMSG, is proposed in this paper to comprehensively capture structural, semantic and attribute information from both homogeneous and heterogeneous neighbors more purposefully. To achieve this, a type-specific linear transformation is first applied to transfer the node attributes to different types of nodes with the same latent factor space. In the new model, the heterogeneous graph is decomposed into multiple metapath-based homogeneous and heterogeneous subgraphs where each subgraph associates specific semantic and structural information; this is different from existing models, which mainly rely on symmetric metapaths. Subsequently, tailored attention-based message aggregation methods are independently applied to each subgraph such that information learning can be more targeted. Finally, information from different subgraphs is fused through graph-level attention to obtain a complete representation. The learned representations are evaluated by several graph-mining tasks. Results indicate that the HMSG attains the best performance in all evaluation metrics than state-of-the-art baselines. Further ablation experiments demonstrate the effectiveness of the modules designed for the HMSG.