Link Prediction In Dynamic Networks Based On Graph Neural Networks

To facilitate the analysis of complex systems,researchers often model them as complex networks,where network nodes represent entities in the system and links represent relationships between entities.These complex systems can be communication networks,transportation networks,etc.Because a lot of useful information can be mined from complex systems,complex network analysis has attracted more and more researchers’ attention.One of the important branches of research in complex network analysis is the Link Prediction,which aims to predict missing links and possible emerging links in a network.Most networks made up of complex systems undergo some structural changes over time,and such networks that evolve over time are often referred to as dynamic networks.Because the topology of dynamic networks changes over time,completing link prediction on dynamic networks becomes a challenging task.If researchers can uncover more hidden information in dynamic networks,the accuracy of link predictions will be further improved.Therefore,in order to solve the problem of dynamic link prediction,this paper proposes two new methods for the characteristics of dynamic networks:(1)Dynamic Link Prediction by Learning the Representation of Node-pair via Graph Neural Networks(DLP-LRN).In this method,a new end-to-end approach is proposed to solve the problem of link prediction on dynamic networks.By utilizing the structural information of a single snapshot,the historical features of network evolution,and the global information of the collapsed network,an improved DGCNN model and an exponential function with learnable parameters are used to effectively learn the embedding representation of target node pairs.Therefore,the DLP-LRN method can be used to deal with and solve the problem of link prediction on dynamic networks.In addition,DLP-LRN is compared with multiple methods on dynamic network datasets,which proves that it has excellent performance in dynamic link prediction.(2)Dynamic Link Prediction by Fusing the History links of Node-pair(DLP-FHN).Considering that previous research did not utilize the link status of target node pairs in historical snapshots,this method mainly considers how to integrate the link status of historical node pairs into some general dynamic network link prediction frameworks.This method proposes three fusion points: first,combining historical link information into a vector and using it as a complementary representation of target node pair embeddings;second,constructing a weighted collapsed graph based on the historical links of node pairs to provide more information for the collapsed graph; finally,this method uses the embeddings of each historical snapshot learned by an improved DGCNN as a new part of the loss function to achieve better training results.To verify the effectiveness of this method,we conducted experiments on multiple datasets.