Deep Learning Based Algorithm For Identifying Critical Nodes Of Complex Networks

With the rapid development of network science and the acceleration of global informatization,complex network theory has provided a practical approach to solving social problems and improving people’s lives.It has been found that very few nodes in the network have an important influence on the structure and function of the whole network.Therefore,identifying the critical nodes in complex networks is significant for theoretical research and practical applications,and its research results have been widely used in virus propagation control,public opinion management,and advertising promotion.However,with the growth of the network scale,the traditional centrality methods cannot well portray the importance of nodes,and the research on identifying critical nodes in complex networks faces new challenges and opportunities.In this context,this paper will identify critical nodes in complex networks based on theories related to graph embedding and deep learning and combining network topology characteristics and node features.The main contents of this research are as follows:(1)In this paper,we propose Inf Res Net,a critical node identification framework for complex networks combining residual network(Res Net)and self-attention.Considering that the influence of nodes and their neighbor nodes in complex networks is mutual,the method first polymerizes each node’s neighbor features in each node’s neighbor network to the node itself and constitutes the node’s feature matrix.Then,the susceptibleinfected-recovered(SIR)model is used to simulate the nodes’ influence,and the nodes with the top 15% of influence are considered critical nodes.The rest are non-critical nodes,which constitute the nodes’ labels.Finally,the feature matrix and labels are used as model inputs to learn node representations by extracting node features through self-attentive and residual networks.The performance of Inf Res Net was evaluated in its entirety across five real networks.The experimental results show that Inf Res Net can effectively identify critical nodes in complex networks compared to the comparison algorithms.(2)In this paper,we propose an RCNN critical node ranking regression model based on the contribution of neighbors N-RCNN.The importance of a node is determined by its neighbor nodes and topology,and the contribution of different neighbors to the importance of a node is different.Therefore,N-RCNN calculates the contribution value of neighbor nodes based on the distance from nodes to neighbor nodes and tight centrality and constructs three feature matrices containing structural information for each node by combining the three centrality features of neighbor nodes.Then,the feature matrix is used as the feature input of the N-RCNN,and the influence of the nodes simulated by the susceptible-infected-recovered model is used as the label of the N-RCNN,and the N-RCNN is trained and predicted.Finally,the prediction results of N-RCNN are ranked.The Kendall correlation coefficients of N-RCNN and six other methods were compared in seven real networks,and the experimental results showed that N-RCNN outperformed the comparison algorithm.