Research On Overlapping Community Detection Algorithm In Complex Networks

With the development of society and information technology,people have found more and more complex systems.In order to get effective information from complex systems,people abstract complex systems into complex networks which are easier to understand.Through research on complex networks,people have discovered that complex networks have universal characteristics such as small-world,scale-free,and community structures.Among them,the community structure in a complex network can be expressed as a set of nodes in the network that have close relationships and edges between them.Community structure's internal nodes are more closely linked to each other and links to other nodes in a complex network are sparse.Due to the characteristics,it can be easily applied to propagation control,demand recommendation,link prediction,and organization architecture analysis.Therefore,how to conduct community discovery efficiently and accurately has become an important research area.In this paper,through the in-depth analysis of the unique problems existing in the two recent community detection algorithms,and corresponding solutions are proposed in order to improve the accuracy community detection.First,when the community detection algorithm based on the Community-Affiliation Graph model is sampled using the Markov Chain Monte Carlo,the high rejection rate of the sampled sample leads to low efficiency of the algorithm and the low probability of detection of the global optimal solution.The advanced Markov Chain Monte Carlo method improves the efficiency and stability of the algorithm.The unadvanced algorithm is prone to "stagnation" during the iterative calculation of the MCMC method.That is,the sampling results obtained by multiple iterations are all the same samples,resulting in uneven distribution of the resulting sample set in the solution space,thereby affecting algorithm efficiency and detection probability of the global optimal solution.Based on this,this paper proposes to introduce a simulation tempering strategy in the MCMC sampling method,adding a second acceptance attempt to the rejected sample,thereby increasing the sampling acceptance probability and reducing the high rejection rate in the sampling process,in order to improve the detection probability and speed of the global optimal solution.Finally,the experimental verification of the four real network data provided by the Stanford University Network Analysis Group shows that the improved scheme is feasible.Second,in the process of finding the community membership matrix by non-negative matrix factorization for the overlapping community detection algorithm based on the Cluster Affiliation Model for Big Networks,the problem is that the local structure around the node itself will affect the membership matrix is ignored.This paper is intended to solve the community problem.When subordinate to the matrix,the local structure around the node itself is fully considered and the influence of the local structure will be embodied by adopting the degree of similarity between the nodes.Based on this,this paper adopts four kinds of inter-nodal similarity algorithms that achieve good results in the field of link prediction and integrate them into the original algorithm.Finally,experimental verification was carried out on four real network data provided by the Stanford University Network Analysis Group.The results show that the improved algorithm improves community discovery quality.