Computational Methods for Learning and Inference on Dynamic Networks

Networks are ubiquitous in science, serving as a natural representation for many complex physical, biological, and social phenomena. Significant efforts have been dedicated to analyzing such network representations to reveal their structure and provide some insight towards the phenomena of interest. Computational methods for analyzing networks have typically been designed for static networks, which cannot capture the time-varying nature of many complex phenomena. In this dissertation, I propose new computational methods for machine learning and statistical inference on dynamic networks with time-evolving structures. Specifically, I develop methods for visualization, tracking, and prediction of dynamic networks. The proposed methods take advantage of the dynamic nature of the network by intelligently combining observations at multiple time steps. This involves the development of novel statistical models and state-space representations of dynamic networks. Using the methods proposed in this dissertation, I identify long-term trends and structural changes in a variety of dynamic network data sets including a social network of spammers and a network of physical proximity among employees and students at a university campus.