Paper detail

Extracting Hidden Groups and their Structure from Streaming Interaction Data

When actors in a social network interact, it usually means they have some general goal towards which they are collaborating. This could be a research collaboration in a company or a foursome planning a golf game. We call such groups \emph{planning groups}. In many social contexts, it might be possible to observe the \emph{dyadic interactions} between actors, even if the actors do not explicitly declare what groups they belong too. When groups are not explicitly declared, we call them \emph{hidden groups}. Our particular focus is hidden planning groups. By virtue of their need to further their goal, the actors within such groups must interact in a manner which differentiates their communications from random background communications. In such a case, one can infer (from these interactions) the composition and structure of the hidden planning groups. We formulate the problem of hidden group discovery from streaming interaction data, and we propose efficient algorithms for identifying the hidden group structures by isolating the hidden group's non-random, planning-related, communications from the random background communications. We validate our algorithms on real data (the Enron email corpus and Blog communication data). Analysis of the results reveals that our algorithms extract meaningful hidden group structures.