|
Abstract:
|
Dynamic networks are used to explore and analyze temporal changes and fluctuations in complex relational systems. As dynamic networks are inherently high-dimensional objects, it is difficult to efficiently capture network dependencies both marginally and through time. In light of this challenge, we propose instead to directly model a low-dimensional representation of the network. We propose and investigate a two-stage modeling procedure that adapts multivariate time series modeling to a family of sufficient statistics that describe the geometric structure of the network. In particular, we consider dynamic networks with temporal heterogeneity under which significant structural changes occur (regime switches) over time. By modeling the sufficient statistics with regime switching time series technique we provide a powerful and interpretable generative model of dynamic networks. We demonstrate the utility of our model through an application to the U.S. Senate co-voting network, which describes the co-voting patterns of members of the U.S. Senate from 1857 to 2015.
|
