Community Detection Based On Graph Neural Network

Community detection as a basic subject in complex networks has been widely used in many fields.Community detection is to mine the community structure of the network by analyzing the topological structure of the network and other information,which is of great significance to understand the network structure and analyze the characteristics of the network.In the early stage,a series of community detection algorithms based on different ideas were proposed,such as label propagation,modularity optimization and so on.The essence of these algorithms is to find the community structure of the network by fitting the topological structure of the network from the perspective of statistics.In recent years,with the rapid development of machine learning and deep learning,community detection algorithm has also been changed from statistical algorithms to machine learning and deep learning-based algorithms,a lot of neural network model such as the feed-forward neural network model,the convolutional neural network model,generative adversarial network model has been applied in the community detection,and has obtained the remarkable effect.Compared with traditional algorithms,algorithms based on machine learning and deep learning can mine the relationship between nodes in the network based on a large amount of data and the powerful fitting ability of the neural network,and then accurately detect the community structure of the network.In recent years,researchers have extended convolutional neural networks to complex networks and proposed a series of graph convolutional network models.As soon as Graph Convolutional Network(GCN)model is proposed,it has been widely concerned by researchers because of its simplicity and efficiency and has been applied in many fields,including community detection.Compared with other community detection algorithms based on deep learning,the community detection algorithm based on graph convolutional network can better fit the topological relationships between nodes in the network.However,the existing community detection algorithms based on graph convolutional network still face some limitations in the accuracy of community detection,mainly because the optimization objective of these algorithms can not represent the community structure well.To address this issue,this paper proposes an overlapping community detection method based on graph convolutional network,from the perspective of enhancing the Markov stability of community structure,which aims to generate community structure with the highest stability through the powerful neighborhood information aggregation ability of graph convolutional network model.Compared with the existing community detection algorithms based on graph convolutional networks,the Markov stability used in the proposed algorithm can better characterize the community structure of the network.In order to demonstrate the effectiveness of our proposed algorithm,we selected two measures to evaluate the performance of community detection,and a series of networks with different scales,including small-scale real-world networks,medium-scale real-world networks,and medium-scale attributed networks as well as the synthetic networks.Experimental results show that compared with other baseline algorithms,our algorithm has the highest accuracy on the majority of these networks.On the other hand,because Markov stability relies on a parameter,Markov time,we found that the accuracy of our method can be improved by optimizing the Markov time in the experiment.Experimental results confirm that given the optimal Markov time,the accuracy of our method can be significantly improved.On the other hand,some community detection algorithms based on modularity optimization have local optimization problems due to greedy policy.In this paper,we propose a community detection method that uses a graph convolutional neural network and reinforcement learning to solve this problem,where the graph neural network is deployed to extract the state of the network,and the reinforcement learning algorithm,deep Q network,is utilized to overcome the local optimal problem in modularity optimization.Experimental results show that the proposed algorithm can avoid the problem of local optimization of modularity,and our algorithm has higher accuracy compared to the existing modularity optimization-based method.