| In complex networks,nodes are usually used to represent entities,and connected edges are used to represent the relationship between entities.Networks are common structures for the representation and analysis of data.Network analysis has important implications for understanding the internal mechanisms of complex systems.After years of research,scholars have found that community structure is one of the common characteristics of complex networks,namely that most of the networks have community structure.Community detection aims to find possible relevant connections from complex networks,mining community structures in the network to help analyze the network.As an important part of network analysis,community detection attracts the attention of researchers in various industries.However,with the increasing complexity of network structure data,the traditional community detection methods have appeared many limitations,including performance and efficiency,and it becomes more and more important to improve the existing community detection algorithms and discover new forms of community detection algorithms.With the development of deep learning,neural networks are also applied to networks.So,a new research direction is born which named graph neural networks.Graph neural network is the expansion of convolutional neural network on graph structure data,which makes up for the inability of convolutional neural network to act on graph structure data,so that convolution can also be used on graph structure data.Community detection is an important application of graph neural networks.In this paper,we study the community detection with graph neural network from two aspects: overlapping community detection and non-overlapping community detection.We conclude that a basic framework of graph neural network applied to community detection: design GNN model according to the task,and then classify or cluster the node features after extraction,so as to obtain the community division of the network.At present,the main GNN definition can be divided into spatial and spectral domain.Spectral domain graph convolution builds the graph neural network model by defining the graph convolutional operators,whose core is the definition of the graph convolution.The Bern Net model uses a Bernstein polynomial approximate filter of order K to design a filter that can learn or design arbitrary filters.However,it is slower than the other methods due to its quadratic dependence on K.Based on this,this paper uses a first-order approximation of Bernstein to define the filter,and designs and implements a supervised community detection algorithm based on Bern Net.Based on the GNN community detection step,the algorithm adopts the first-order Bernstein polynomial approximation to define the filter,takes the original label of the data as the basis for the community division,define the loss function between the label predicted by the model and the original data label,learned the probability model of the node label predicted by the node feature matrix,and classify the predicted node label probability,thus obtain the final community division.Experimental results show that the algorithm can achieve better community detection results on real datasets and reduce the algorithm complexity while obtaining better results.To study the problem of how the graph neural network is applied to the community detection task,the main work in this paper is summarized as follows:First,we introduce the application of deep learning to community detection problems,expounds the basic framework of community detection in graph neural networks and provides idea for the research of this paper.Second,an appropriate GNN definition was chosen to construct the GNN community detection model.Bern Net is a graph convolutional neural network based on a Bernstein polynomial of order K approximation that can approximate or design arbitrary filters,but with slower computation due to its dependence on K.In order to speed up the calculation speed and reduce the computing cost,this paper adopts the Bernstein first-order polynomial approximation definition filter to improve this and propose a supervised community detection algorithm based on Bern Net.Finally,the algorithm is applied to the real data set for community detection to obtain the corresponding community division results,and the standardized mutual information(NMI)is used to evaluate the community division results,and compare many different community detection algorithms to make the experimental data more convincing.The experimental results show that the Bern Net-based community detection algorithm can obtain good community division results and greatly improve the operation speed. |
