Link Prediction In Heterogeneous Networks Based On Graph Neural Networks

Most systems in the real world can be modeled as complex networks,and with the development of big data in networks,analyzing complex networks can reveal more potential valuable information from various systems.Among them,link prediction is one of the key issues in the analysis of complex network information,which aims to predict or determine whether there exists or will exist a connection between two nodes by analyzing the relationships between nodes in the network.Link prediction can not only discover missing links and predict the appearance of new links in the network,but also help researchers understand the global structural characteristics and evolution mechanisms of the network.By analyzing the linking patterns between nodes in the network,information such as community structure,node influence,and key nodes can be revealed.Currently,node embedding methods based on graph neural networks have achieved superior performance in link prediction tasks,but due to the low-pass filtering characteristics of GNN,these methods perform poorly in highly heterophily networks,and there is less research on link prediction in heterogeneous networks that contain multiple types of nodes and edges.In heterogeneous networks,different types of nodes have different feature space distributions,so link prediction methods for homogeneous networks cannot be directly applied to heterogeneous networks.In order to solve the link prediction problems in heterogeneous and heterogeneous networks,we propose two end-to-end link prediction methods that are suitable for heterophily and heterogeneous networks,respectively.(1)For the task of predicting links in heterophily networks,we propose a link prediction method based on reconstruction network in heterophily networks,called LPHNR.First,this method extracts closed subgraphs according to target node pairs,and measures the heterophily degree of the target node pairs by calculating the heterophily coefficient of the entire subgraph.Then,neighboring nodes are labeled based on the heterophily coefficient of the node pairs and the similarity of label attributes,and beneficial,redundant,or harmful node information is selected for downstream link prediction tasks.Next,node pairs are labeled based on their heterophily coefficient,embedding similarity,and the types of neighbor node labels.If a node is labeled as high-quality beneficial information,a link is added between the node pair if there is no link between them;if the node is labeled as redundant information,the edge is not modified;if the node is labeled as harmful information,a link between the node pair is removed if there is one.Finally,the SEAL framework is used for embedding learning on the reconstructed network.Experimental results show that LPHNR has superior prediction performance on both homophily and heterophily networks.(2)For the task of link prediction in heterogeneous networks,we propose a heterogeneous network link prediction method called LPMPA based on meta-path projection.This method learns embeddings of node pairs from different meta-paths through metapath projection and semantic graph aggregation learning.Specifically,we first project the heterogeneous network into multiple isomorphic semantic graphs based on multiple meta-paths.Then,we extract probability subgraphs from each semantic graph and use a GNN to learn the embeddings of node pairs in the subgraphs.Finally,we design a semantic aggregation module that combines the embeddings of node pairs obtained from different semantic graphs using attention mechanisms to obtain the final embeddings.To validate the performance of the proposed method,we conducted experiments on three heterogeneous datasets.The results show that the accuracy of the LPMPA method is better than that of seven baseline methods.