Inferring Polyadic Events With Poisson Tensor Factorization

We present a Bayesian factorization model for discovering dynamic communities of actors in a social network and for describing action types associated with those communities. The structure of a social network is often hidden in the interactions among its actors. Typically, we observe interactions between pairs of actors (i.e., an edge) and our goal is to uncover their more complex relationships. For example, here we consider a problem in international relations where our goal is to analyze pairwise country interactions to infer communities of countries that are organized into international “events.” We develop a Poisson factorization model for decomposing tensors of time-stamped and typed interactions; and we present a mean-field variational method for efficiently fitting this model to large amounts of data. We study the inferred components from a real-world data set of country– country interactions, and show that they conform to and inform our knowledge of international affairs.