Mixed-Membership Stochastic Blockmodels for Bipartite Networks

• In Event: Advances in Probabilistic Network Data Analysis

Abstract

A substantial number of networks of interest in social and political research are bipartite --- i.e. they consist of two distinct types of nodes, and edges that are only defined between (rather than within) node types. Common approaches to the study of these types of networks rely on marginalizing edge information over one of the node types, thus forcing researchers to ignore potentially relevant node-level information and risking the introduction of aggregation bias. To address these issues, we propose an extension of the mixed-membership stochastic blockmodel that respects the bipartite nature of such networks. The model estimates type-specific communities and can incorporate node-level attributes that are associated with membership therein. In this paper, we introduce our model, present a scalable estimation algorithm for it, and illustrate the full approach using data on bill cosponsorship decisions in the US Senate.

Authors

• Adeline Lo, University of Wisconsin, Madison

• Santiago Olivella, University of North Carolina

• Kosuke Imai, Harvard University