Attributed Heterogeneous Graph Embedding with Meta-graph Attention

Attributed heterogeneous graph is widely utilized to model different types of objects and relationships in real world. Due to the capability of graph embedding in learning the effective features of nodes, the methods for attributed heterogeneous graph embedding are proposed to enhance downstream tasks. Existing methods mainly use meta-paths to explore the semantic and structure information in the graph. To capture richer and more meaningful information, we propose a novel model Attributed Heterogeneous graph Embedding with Meta-graph Attention (AHEMA) by exploiting meta-graphs and the attention mechanism. Firstly, the features of neighbors based on a meta-graph are aggregated by discriminating both the differences of neighbors and their types. Then, to enhance the node embeddings, the instance embeddings of a meta-graph are learned and incorporated by constructing an instance graph. Finally, the node embeddings under different meta-graphs are fused by considering the importance of meta-graphs. Experimental results on three real datasets show the proposed AHEMA model outperforms the baselines on node classification and node clustering tasks.