Link Prediction In Multiplex Networks Based On Graph Neural Network

Complex networks can be used as a tool to describe real scenarios,so analyzing and studying complex networks can lead to a better understanding of the structure and dynamics of real-world systems.Nowadays,the rapid development of the Internet has promoted the rapid progress of data mining and other aspects.At the same time,as an important branch of complex networks,link prediction has also attracted more attention of researchers with the improvement of data analysis and computing capabilities.The research objective of link prediction is to explore the unobserved links in a network through known node information and network structure.These links may represent the potential connections within the network or the connections that may form in the future.Therefore,link prediction has significant contributions in improving search experience,optimizing recommendation systems,and understanding user behavior.In a network system,multiplex networks can better describe the relationships that exist in real scenarios,where each layer expresses various interactive dependencies.Numerous studies have indicated that the formation of links in a layer can be influenced by corresponding nodes in other layers.However,how to effectively use the information and structure in multiplex networks to improve the predictive ability of the model remains a challenging task.An approach that takes into account the structure of all layers and extracts effective information from it for link prediction is desirable.At present,numerous solutions have been suggested for addressing link prediction problems,but few people pay attention to using graph neural network to solve link prediction problems in multiplex networks.Hence,this paper proposes a link prediction method using graph neural network in multiplex networks,which considers the information of the whole network and transforms the traditional link prediction into a binary classification problem in which the network can learn independently and filter the information.The main tasks and innovations are as follows:(1)A link prediction algorithm in multiplex networks based on information filtering fusion and graph neural network.Node pairs can enhance their own information by utilizing the neighbor information in the subgraph they select,and studies have found that not all known information is valid.Therefore,this study designed a method to remove redundant information and add potential links,thereby obtaining effective information and features for each corresponding node at each layer.The attention mechanism is used for multiplex network feature fusion,and then the fused features are fed into the DGCNN model for training to obtain the classification results,whether the link exists or not.Compared with other multiplex network link prediction methods,this algorithm fully considers the effectiveness and potential of information,and takes into account the node information and structure information of multiple network layers,so that the information used in this paper in the prediction is reasonable and effective.By analyzing the experimental results of six real networks,the effectiveness of the proposed method has been demonstrated,and its superior performance in terms of execution capability over other multiplex network link prediction methods has been shown.(2)A link prediction algorithm in multiplex networks based on interlayer interaction and graph neural network.The work in the previous section demonstrated the need for information filtering and fusion,but the fusion of information across multiple layers was simply processed using an attention mechanism,which did not guarantee the accuracy of the fused information.Therefore,the work in the second part of this paper considers these problems and improves them by fusing the features of the multiplex network with the information of the trained filtered graph separately to ensure that the learned content is true and accurate.In addition,the study of asymmetric neighbourhood similarity reveals that the influence on the target layer does vary between layers,so the role of layers in the multiplex network is also explored based on the first work,and this work assigns different weights between the newly fused layers to indicate the degree of influence on the target layer.The above two improvements are combined with the first work to propose a link prediction algorithm in multiplex networks based on interlayer interaction and graph neural network.By comparing the results of this method with the results of the comparative methods,it is found that there is a significant improvement in all the metrics,which also justifies the need for the improvements and the conjecture of this study.Through the aforementioned research,this article explores the use of graph neural networks for multiplex network link prediction algorithms and addresses some of the issues in this direction,providing new perspectives for multiplex network link prediction.