Overlapping Community Detection in Social Networks by Local Expansion

Abstract

Community structure is one of the main structural features of networks and detecting overlapped community structure is an important field in social network analysis. There are many methods for finding non-overlapping communities in this research area. The existing studies about overlapping do not sufficiently address the problems of the relationship between objects in overlapping regions and the roles of these objects during the formation and growth of communities. In recent years, local community detection algorithms which detect overlapped community structure have been developed. Local expansion methodologies that detect local community structure are techniques to find a community through the seed. Therefore, recent algorithms have emphasized on the locating seed rather than random seed selection. However, although the most existing algorithms could identify superior seed, their expansion strategies did not become effective and efficient strategies. Moreover, algorithms suffer unstable community structure because the influences of parameter for controlling community’s resolution of fitness evaluation functions where used in community expansion process. In this research, therefore, the algorithm is modelled on local expansion strategy and designs the extended jaccrad similarity to find seed. In addition, this research formulates the optimized parameter evaluation formula to avoid the parameter influences. This work, firstly, identifies the seed or core node by using extended jaccard similarity and form initial community via seed. Then local community is detected by expanding the initial community with fitness function based on proposed optimized parameter evaluation and finally overlapped nodes are identified by merging detected local communities. In this dissertation, the algorithm is implemented by using small datasets from network data repository site and large networks from Stanford large network datasets collection. In addition to real networks, overlapping artificial benchmarks are also selected to generate the experiment networks. On both real and artificial, the performance results of proposed algorithm are compared with state of the art algorithms by using various performance evaluation metrics. In particular, the proposed algorithm is proven that it has better accuracy on both real and benchmarks and saves running time as an efficient algorithm.