Heterogeneous Graphs Embedding Learning with Metapath Instance Contexts

Chengcheng Yu, Lujing Fei, Fangshu Chen, Lin Chen, Jiahui Wang

Abstract: Embedding learning in heterogeneous graphs consisting of multiple types of nodes and edges has received wide attention recently. Heterogeneous graph embedding is to embed a graph into low-dimensional node representations with the goal of facilitating downstream applications, such as node classification. Existing models using metapaths, either ignore the metapath context information which describes intermediate nodes along a metapath instance, or only using simple aggregation methods to encode metapath contexts, such as mean or linear methods. To address the problems, we propose Metapath Instance Contexts based Graph neural Network (MICGNN). Specifically, MICGNN includes four components: Node transformation, projects the input different types of features into the same feature space. Context aggregation, incorporates the node embeddings along a metapath instance context. Instance aggregation, combines context embeddings obtained from each metapath instance. Semantic aggregation, fuses together the semantic node embeddings derived from different metapaths. Extensive experiments results on two real-world datasets show our model exhibits superior performance in node-related tasks compared to existing baseline models.