Research On Community Detection Algorithm Based On Graph Representation Learning

With the development of science and technology,the network has penetrated into all aspects of people’s life,such as social network,citation network,biological network and so on.Many things in life can constitute a kind of network data.There is a kind of cluster structure in the network,which we call the community.Community structure can reflect the internal characteristics and functions of the network.Discovering the community structure in the network is helpful to understand the hidden information in the network,from which we can find the internal changes of the network.Capturing such changes is beneficial to grasp the evolution trend of the network.Let the network get more efficient application.Traditional community detection methods are mostly based on statistical reasoning and traditional machine learning.With the continuous expansion of network size,such methods are no longer sufficient to cope with the more complex data and social scenarios.Graph representation learning uses deep learning technology to transform high-dimensional network nodes into vector representations of low-dimensional space,while maintaining network structure information and attribute information,and then applies it to subsequent graphical tasks,such as connection prediction,node classification,etc.It can learn more complex graphical representations.The community detection algorithm based on graph representation learning firstly uses graph representation to embed the network into low-dimensional vector space,and then uses clustering algorithm to cluster node vectors into community structures.However,However,traditional clustering algorithm cannot effectively cluster node vector at community level,makes the community of clustering results is difficult to reflect the community characteristics of "high cohesion and low coupling",and now most of graph representation learning is not designed for community detection task,not geared to the needs of the community level clustering optimization,the output node vector does not apply to the community detection task.Aiming at the above two problems,this paper improves the community detection algorithm based on graph representation learning from two aspects of clustering and graph representation learning respectively.The main research contents and innovation points are as follows:(1)In order to solve the problem of the general clustering algorithm cannot reflect the community feature in the clustering of graph represented node vectors,this paper offers a community detection model based on clustering cover algorithm(CCL-CD).The clustering cover algorithm firstly obtains the head node according to the point density in the graph representation feature space,adjusts the position of the head node,and takes the convergent head node as the cover center.Then each node in the network is given an initial label according to its distance from the cover center,and the average distance between similar nodes and the cover center is taken as the cover radius to form a spherical cover.After the cover is formed,the community structure is obtained by the secondary division of the uncovered nodes combined with the network structure information.In this paper,we combine the clustering cover algorithm with graph representation learning algorithm to deal with the community detection task,and it is shown that the proposed algorithm can obtain more reasonable community results in several experiments with and without real label networks.(2)In order to solve the problem of the current graph representation algorithm only considers the network Structure and Properties,but does not consider the optimization of clustering at the community level,this paper offers a community detection model based on graph representation learning with clustering information(GRC).GRC model is divided into three parts: graph representation module,graph generation module and clustering module,graph representation module and graph generation module to form a encoder,in the form of unsupervised learning node vectors,and the node vector as a community membership matrix,do not rely on clustering algorithm for community structure can handle more diverse community relations.Based on the idea of deep clustering,the clustering module calculates the clustering loss of node vectors to optimize the graph representation model parameters,generates the node representation with clustering information,and realizes the optimization of the graph representation algorithm at the community level.In the comparative experiment,this paper compares the GRC model with the overlapping community discovery algorithm and the nonoverlapping community discovery algorithm on the real network,and verifies the effectiveness and feasibility of the model.