Discovering and Exploring Overlapping Community Structures in Large Networks

Опубликовано 21 июня 2016, 18:43

Networks are ubiquitous in our life. Examples include social networks, computer networks, and biological networks, among others. In this talk, I will present a novel scalable approach to addressing a fundamental problem in network analysis: how to effectively detect overlapping community structures in a large-scale network so that the subsets of nodes within the same community tend to share similar properties? We build our approach on a new triangular characterization of networks and a fast stochastic variational inference (SVI) algorithm, yielding an efficient inferential procedure that scales linearly in both the number of nodes and the number of communities. Empirical results show that our triangular model SVI procedure is not only faster but also more accurate in terms of community recovery on large networks. We also demonstrate that our method is able to discover interesting communities on a massive IMDB co-actor network with 896K actors.