Link Prediction In Heterogeneous Networks Based On Neighbor Aggregation

Graph data mining typically involves abstracting relationship information from reality to generate complex graph data to mathematically characterize real-world problems.As an important branch of graph data mining,link prediction has significant implications in fields such as friend recommendation,relationship inference,and evaluation prediction.Heterogeneous networks,as complex networks with different types of edges or nodes,can provide more detailed descriptions of complex situations in the real world.For the same reason,scholars have proposed numerous algorithms to solve the problem of link prediction in heterogeneous networks.Among them,the heterogeneous network link prediction method based on meta-paths and graph neural networks has been widely recognized by scholars.This paper aims to study the problem of link prediction in heterogeneous networks and proposes two algorithms: the type decomposition link prediction method based on meta-paths and the link prediction method based on heterogeneous line graph neural networks.(1)The type decomposition link prediction method based on meta-paths.Existing heterogeneous network link prediction methods usually aggregate neighbor nodes based on meta-paths and consider them equally important,and then aggregate them between different meta-paths.Obviously,standardizing the aggregation within the meta-path is not detailed enough.Therefore,this method improves the aggregation method by targeting specific types of nodes within the meta-path,in order to more finely distinguish the contributions of different types of nodes to the meta-path.As the contributions of different types within the meta-path are refined,our algorithm achieves an improvement in prediction accuracy.Experimental results show that compared to the other nine baseline methods,our model has better accuracy and stability in the selected network.(2)The link prediction method based on heterogeneous line graph neural networks.Due to the problem of manually selecting meta-paths in existing heterogeneous network link prediction methods,we choose to solve the problem by line graph conversion.By converting the heterogeneous graph to a heterogeneous line graph,more information is retained while reducing the difficulty of processing the heterogeneous network.On this basis,we introduce the idea of handling heterogeneity in isomorphic networks to heterogeneous networks during aggregation,and propose a new aggregation method based on transforming node types in the transformed heterogeneous line graph.In this way,we can better distinguish the contributions of different types of nodes in the heterogeneous line graph.We conducted extensive experiments to evaluate the predictive performance of the proposed model.Experimental results show that compared to the other ten baseline methods,our proposed model achieves better link prediction accuracy.Overall,this article focuses on researching the link prediction problem in heterogeneous networks and proposes two different approaches for link prediction algorithms.These algorithms are tested on several common datasets.