Research On Link Prediction For Multiplex Networks

With the advancement of science and technology and the development of human society,various complex systems are mushrooming in today’s world.Complex net-works are effective way to model complex systems,hence complex network analysis has attracted the research interest of more and more scholars from different disciplines.Link prediction is an important research branch of complex network analysis,which aims to discover lost links in the networks and predict future links in the networks,and can also infer spurious links existing in the networks.Therefore,it has important theo-retical significance and great application value in the real world,and has always been a research hotspot in recent years.As a special type of network in complex networks,multiplex networks can describe different types of relationships between the same set of entities.Compared with the single-layer networks,the modeling ability of the multiplex networks is more powerful and more realistic.Therefore,this dissertation focuses on link prediction in multiplex networks.Among the various methods of link prediction,the similarity-based methods are simple,interpretable and have high prediction accuracy.Therefore,two similarity-based link prediction methods in multiplex networks are proposed successively using multi-attribute decision-making and evidence theory in this dissertation.In addition,because graph neural networks exhibit excellent performance in processing unstructured data,the link prediction problem in dynamic multiplex networks based on graph neural network technology is studied in this dissertation.(1)In order to accurately predict links in multiplex networks,a novel method for link prediction in multiplex networks based on multi-attribute decision-making is pro-posed in this dissertation.The structural features of different layers in a multiplex net-work are interrelated to a certain extent.Therefore,effectively utilizing the information of different layers can improve the accuracy of link prediction.This dissertation consid-ers link prediction in multiplex networks as a multiple-attribute decision-making prob-lem,in which alternatives are potential links in the target layer and attributes are diverse layers in the network.To this end,a novel multiple-attribute decision-making approach is proposed to fuse the structural information of all layers of the target node pairs in this dissertation.To weight each layer in the proposed method,a layer similarity measure is defined based on cosine similarity.Experimental results show that the proposed method has better performance than competing methods in terms of both accuracy and running time.(2)In order to further improve the accuracy of link prediction in multiplex net-works and predict the links in negatively correlated multiplex networks,a novel method for link prediction in multiplex networks based on evidence theory is proposed in this dissertation.The proposed method gauges the connection likelihood of a node pair by integrating its similarity scores from all layers using evidence theory.In the proposed method,each layer is regarded as a source of evidence,and the similarity of each node pair in one layer is represented by a mass function.To calculate the similarity scores of node pairs more precisely,a novel basic similarity index is defined inspired by the conductance model.In addition,we evaluate the reliability of each evidence source by calculating the correlation of the other layers with the target layer,and discount ev-ery piece of evidence based on the reliability of the evidence sources.Furthermore,the proposed method is partially improved to reasonably conduct link prediction in the negatively correlated multiplex networks.To analyze the effectiveness of the proposed method,extensive experiments were conducted over eight positively correlated mul-tiplex networks and four negatively correlated multiplex networks,and experimental results show that the proposed method outperforms existing methods in most cases.(3)Aiming at the low accuracy of current link prediction methods in dynamic mul-tiplex networks,a novel method for link prediction in dynamic multiplex networks based on graph neural network technology is proposed in this dissertation.The proposed method extracts the subgraph for each target node pair in the snapshot of the current moment of the target layer.This method takes into account the dynamic properties of the dynamic multiplex networks.When constructing the feature matrix of one target node pair,in addition to using the labels and embeddings of the nodes in the subgraph,the method also considers the information in the historical snapshots of the target node pair,so that the proposed method can fully extract the required dynamic information.This method also considers that the dynamic multiplex networks have multi-layer prop-erties,and encodes the information in the auxiliary layers of the target node pair,thus effectively utilizing the useful information provided by the auxiliary layers.Then,all the extracted effective information is received for training and the prediction results are outputted by Deep Graph Convolutional Neural Network(DGCNN).The experimental results prove that the proposed method outperforms multiple types of baseline methods and has excellent performance.