New developments in Exponential Random Graph Modeling
Razib, Raihan Habib
MetadataShow full item record
This dissertation offers research contributions pertaining to the use of Exponential Random Graph Modeling (ERGM) for social network analysis in terms of application of observational social network, new algorithm development and use of ERGM in virally disseminated surveys. The first research thrust develops a methodology for evaluating the strength of cause-and-effect relationships in networks. Given the evidence of peer pressure impacting human decision-making and the fact that growing online social networks now make us more connected than ever, this is a timely contribution. Separating the challenges of 'dependent self-selection' and 'interference' in network settings. I present an approach to quantifying the effects of social network dependencies on self-selection by advancing the theories and applications of two research areas - Exponential Random Graph Modeling and "propensity score" based observational causal inference. The second research thrust presents a new Markov Chain Monte Carlo-type method for sampling from Exponential Random Graph Models efficiently with the help of Balance Subgraph theory, which is a special branch Random Graph theory. We show that this new sampling technique can help us approximate complex network sufficient statistics in constant time which is impossible for the currently employed exhaustive enumeration based approach. The presented method is applied for inferring Exponential Random Graph Model parameters from an array of diverse social networks. Its performance is found to favorably compare to the conventional Gibbs sampler especially on sparse networks. The third research thrust presents the idea of using Exponential Random Graph Models to analyze virally disseminated survey data (e.g., online survey, in a viral manner). This research also shows how ERGM can be used for advertising by inferring response of non-respondent.