Heterogeneous Graph Neural Network Focused on Structural Features for Link Prediction

Heterogeneous Graph Neural Networks (HGNNs) inherit some of the mechanisms of traditional graph neural networks, and are able to focus on graph structures that contain different types of nodes and edges, effectively embedding their structural and semantic features into node representations. Little work has investigated whether these capabilities play a critical utility in the problem of link prediction for heterogeneous graphs. In this paper, we target these issues by proposing Heterogeneous Graph Neural Network Focused on Structural Features (SFH4LP). Mean value aggregation is used to simplify the computation of neighbor aggregation, while centrality coding is added to complement node representation. In order to exploit the structural information of the graph more comprehensively, the model pays deep attention to the edges and spatial relationships in the metapaths. An improved self-attention mechanism based on Transformer is proposed to explore the information implied on the metapaths and achieve high-quality fusion of features from different metapaths to generate the final node representations. Experimental results on three heterogeneous graph datasets show that the model proposed in this paper outperforms existing baseline models in terms of the effectiveness of heterogeneous graph embedding, and exhibits good performance for the link prediction task.