Seminar by Peter MacDonald

Please Note: This seminar will be given in person.

Structured latent space models for multiplex network data

Network data, made up of pairwise connections (edges) between a set of nodes, appear in many important statistical applications, including neuroimaging, sociology, and economics. These data are often collected through the observation of highly complex and interconnected systems and are naturally represented as multiplex networks: multiple network layers, each encoding a different type of edge for a common collection of nodes. Models and methods in statistical network analysis are traditionally designed for a single large network, but applying these methods to an aggregated multiplex network can miss important heterogeneity in the data. Our new methodology instead provides a comprehensive framework for modeling, estimating, and summarizing multiplex networks which decomposes the structural similarities and differences between layers in a data-driven way. I will also briefly discuss a second model and data-driven smoothing approach for multiplex networks with naturally ordered layers, for instance the order inherited from layer-specific time stamps. In both settings, we parameterize the multiplex network through structured, low-dimensional latent positions, which govern the activity of each node on each layer. We provide adaptive estimation algorithms, prove theoretical results for recovery of unknown parameters, and apply our methods to simulated and real data, including multiplex networks of international trade, and diplomatic interactions.