Statistical Modeling of Large-scale Network Data

Abstract

The explosive growth of network data in various domains has spurred extensive research on statistical modeling techniques to unravel the underlying patterns of complex interaction processes. This thesis addresses the challenges and opportunities presented by large-scale network data. In particular, the focus is on the development of novel statistical models to gain insights into community structures and predict future links. The research comprises three key projects, namely community detection within edge exchangeable model, community detection in the scenario where interactions exhibit edge clustering, and conformal link prediction. This work advances the existing research in several directions. First, it extends conventional node-centric models for network analysis to interaction process-based models using the concept of edge exchangeability. Second, scalable algorithms are implemented and applied to real-world data that are composed of millions of interactions. Third, it transforms the elucidation of the network structures into a predictive tool for future interactions using conformal prediction. Throughout this thesis, extensive experiments are conducted on a series of simulation data and real-world datasets, such as TalkLife and Enron data, to evaluate the performance and effectiveness of the proposed methodologies. The results demonstrate the superiority of the developed statistical models in capturing the complexities of large-scale network data and providing valuable insights into community structures and link prediction.