Forecasting Dynamic Group Behavior in Social Media

We propose to develop novel, scalable technology for detection and of tracking of new and existing groups in social media, together with a comprehensive set of indicators for inferring group activities and interactions. Focusing initially on Twitter data, our approach will extend existing"static"models of community structure to properly fuse both topic-based and topological information, and will use a Markov chain Dirichlet process to model the evolution of this structure over time. We will also develop on-line variational methods for on-line group detection and tracking at web scale using this model. Finally, we will employ Bayesian methods and Granger causality to develop indicators of activities within and between groups, including recruitment and competition, and we also will use these indicators to improve group detection and tracking.