Statistical Learning and Dynamic Networks


Our research focuses on scalable approaches for modeling and understanding properties of large dynamic networks. In particular, we are developing statistical learning methods for describing the evolution of networks where both node attributes and link topology can evolve under mutual influence. We are also developing novel information-theoretic approaches for detecting causal relationships in dynamic network that cannot be detected by analyzing the network topology, and examining information flow in such dynamic networks.

Our work addresses the impact of network structure on its functionality, and borrows from statistical physics of disordered systems to examine fundamental limits on one's ability to detect such structures. For instance, it is known that in unsupervised community detection there is a threshold of the community-overlap beyond which communities cannot be detected. From the perspective of statistical inference, this type of phase transition between detectable and undetectable regimes is undesirable, as it signals inference instabilities - large fluctuations in accuracy in response to relatively small shifts in the parameters. We have demonstrated that any generic amount of prior information, however small, destroys the critical nature of the inference, and shifts the community detection threshold its lowest possible value.