Community detection in dynamic networks

A reasonable representation of some complex systems such as social and biological systems is a network topology that allows its components and interactions among them to change over time. Understanding the time-dependence of these networks can lead to invaluable insight about characteristics and structure of time-varying networks. In this thesis, several classes of static and dynamic clustering algorithms and ideas are reviewed. A challenge arising in dynamic clustering schemes is that the detected communities are not independent over time and the identified clusters at one point of time should not dramatically deviate from the results of previous timesteps. It is especially important to reduce large short term variations and ensure that communities smoothly change over time. Here we present a novel method which is built upon a probabilistic generative Bayesian model to address the problem of identifying consistent and stable overlapping communities in dynamic networks. Synthetic and real networks are used to evaluate the performance with respect to different parameter settings, the model order selection, and the run-time of the proposed algorithm. Performance analysis indicates thatthe algorithm proposed in this thesis outperforms several other state-of-the-art algorithms and provides valuable insights into the evolution and underlying structure.