Dynamic Community Detection with Temporal Dirichlet Process

Research of detecting dynamic communities from network stream has attracted increasingly attention recently. Some of the previous techniques employed a two-stage approach to detect communities. However, since the two-stage approaches detect communities within each epoch independently, the identified communities usually have high temporal variation. Another restriction of the previous techniques is the requirement of predefining the number of hidden communities by a fixed value or within a very narrow range. To overcome these limitations, we propose the Dynamic Stochastic Block model with Temporal Dirichlet Process, which is able to detect communities and track their evolution simultaneously from a network stream. The number of communities is automatically decided by a Recurrent Chinese Restaurant Process without human intervention. In addition, the identified communities exhibit a rich-gets-richer effect and other appealing properties. The experiment results on both simulated dataset and Flickr dataset showed the effectiveness of our proposed technique.