A Network Model for Dynamic Textual Communications with Application to Government Email Corpora

In this paper, we introduce the interaction-partitioned topic model
(IPTM)---a probabilistic model of who communicates with whom about
what, and when. Broadly speaking, the IPTM partitions time-stamped
textual communications, such as emails, according to both the network
dynamics that they reflect and their content. To define the IPTM, we
integrate a dynamic version of the exponential random graph model---a
generative model for ties that tend toward structural features such as
triangles---and latent Dirichlet allocation---a generative model for
topic-based content. The IPTM assigns each topic to an "interaction
pattern"---a generative process for ties that is governed by a set of
dynamic network features. Each communication is then modeled as a
mixture of topics and their corresponding interaction patterns. We use
the IPTM to analyze emails sent between department managers in two
county governments in North Carolina; one of these email corpora
covers the Outer Banks during the time period surrounding Hurricane
Sandy. Via this application, we demonstrate that the IPTM is effective
at predicting and explaining continuous-time textual communications.