Research On Network Community Detection Algorithm Based On Dynamic Model

With the rapid development of science and technology,human society is constantly moving forward.All kinds of things in society and the relationship between people are becoming more and more complex,forming a large-scale,complex network system.This series of phenomena prompt the study of problems related to complex network to become a hot spot in the current era.Among these problems related to complex networks,the problem of community detection is one of the most important branches.This problem is also an important substep in the process of data mining and knowledge discovery.Detecting the community structure in the network is of great significance for both revealing the essential information in the network and deeper research on the relationship between things.Due to the universality of the community structure,the problem of community detection has penetrated into many subjects such as physics,biology and sociology.The research of community detection has stimulated the interests of experts from various fields.In this paper,we propose two methods for community detection from the perspective of dynamics.The first community detection method is on the base of distance dynamic synchronization,which is abbreviated as DDS.The method first uses the Jaccard distance as the initial distance of each edge in the network,and each vertex is considered as a community.Then according to the inherent topology of the network,DDS continuously updates the distance of each edge in the network in an iterative manner until the distance of each edge finally reaches a steady state.In each distance update process,we first perform merging and division procedure according to the distance in the current iteration,and then we calculate the influence to the distance from the neighbors of the two endpoints by using the connection relationship between vertexes.The distance between vertexes belonging to the same community will shrink,while distance between vertexes belonging to different communities will stretch.The second method,which is based on similarity variance,is an improvement of the traditional label propagation algorithm,abbreviated as SVLPA.The SVLPA algorithm first calculates the similarity variance of each vertex,and sorts the vertexes in descending order of similarity variance values as the vertex label update order.When a vertex in a network updates its own label,we select the label of the neighbor vertex with the highest degree of similarity to the current vertex as the current vertex label among the neighbor vertices with the most frequently occurring labels.By taking these two strategies,SVLPA overcomes the disadvantage of unstable community results generated by traditional label propagation algorithm on the premise of ensuring the approximate linear time complexity.Finally,we evaluate the two community detection algorithms proposed in this paper on a large number of real datasets and artificial datasets.The experimental results show that high quality community structure can be spotted without any human intervention under the two schemes in this paper.DDS accurately displays the community structure in the network in an intuitive manner,and at the same time it can effectively avoid the "extreme" distribution of the community.SVLPA detects stable and accurate community structure from the network under the premise of ensuring the time efficiency of the original label propagation algorithm.Experiments on the Ring network further illustrate that the two proposed algorithms in this article can avoid the effect generated by "resolution limit" problem.