A dynamic clustering based method in community detection

Abstract

Social networks are growing, community detection has become one of the hot topics in social network research. As various types of social networks continue to emerge, universal community detection approaches are becoming increasingly important. As the real community division is dynamic, the community structure will appear or disappear with the passage of time. Therefore, the authenticity and real-time of the division become the core foundation of community detection, and the design of a real-time algorithm based on real division has great challenges. In this paper, we propose a dynamic community detection algorithm dynamic clustering by fast search and find of density peaks (D-CFSFDP) based on the partition of nodes-follow relationships to improve the accuracy and adaptability of real complex community detection. In D-CFSFDP, a distance metric based on trust is defined, the user relationship in the social network is quantified as a distance matrix, and the size of the matrix element is used to measure the degree of the user relationship. Then we use kernel density estimation on the distance matrix, and compile the statistics of the impact of each node in the network. We combine the improved KD-Tree model and mean integrated squared error criterion to improve the calculation flow, so that it adapts to different sizes of data sets to improve the calculation accuracy. Based on the principle of density peak clustering and the community attributes, the internal structure and natural outside structure of the community can be obtained according to the distance between the nodes. Finally, the remaining nodes are allocated by distance to the corresponding community to complete the community division. The static community division is further extended to a dynamic detection algorithm that gets linear time complexity. Therefore, we can change the community structure by updating the node relationships in the network. Through the visualization software we can observed that, the D-CFSFDP algorithm give the results of community division with a clear natural and internal hierarchical structure. With the increase of community scale and difficulty of division, D-CFSFDP algorithm has excellent stability. In the real data set and the Douban network, the community division is more close to the real division result, the adaptability is good and the feasibility and validity are verified.