Title:Using Model-based Overlapping Seed Expansion to detect highly overlapping community structure

Abstract:As research into community finding in social networks progresses, there is a need for algorithms capable of detecting overlapping community structure. Many algorithms have been proposed in recent years that are capable of assigning each node to more than a single community. The performance of these algorithms tends to degrade when the ground-truth contains a more highly overlapping community structure, with nodes assigned to more than two communities. Such highly overlapping structure is likely to exist in many social networks, such as Facebook friendship networks. In this paper we present a scalable algorithm, MOSES, based on a statistical model of community structure, which is capable of detecting highly overlapping community structure, especially when there is variance in the number of communities each node is in. In evaluation on synthetic data MOSES is found to be superior to existing algorithms, especially at high levels of overlap. We demonstrate MOSES on real social network data by analyzing the networks of friendship links between students of five US universities.