In this talk I will introduce a framework for modeling social network interaction data using Bayesian Networks and briefly discuss an algorithm that can learn the Bayes Net structure learning for this case efficiently. This algorithm allowed us to learn Bayes Nets for as many as 3 million people with a node per person. I will then show how to incorporate additional information about people, such as their interests and affiliations by introducing a generative model and a corresponding scoring metric that combine a block modeling approach with Bayes Net structure learning. Experimental results show improvement on several fronts: the Bayesian Networks overfit less when additional information is used, and learned latent clusterings of people help to provide meaningful insights into social groupings.
This work was done jointly with Andrew Moore and Zoubin Ghahramani and is based on my thesis (Goldenberg, 2007).