A Path Weighted Regression Model for Big Network Data

• In Event: Structure and Change in Strategic Communication Networks

Abstract

Research in Political Science is now benefiting from massive data sets reporting inter-connected observations. Social media data provides one of the most visible examples, but is hardly unique. Modeling Big Relational Data (BRD), however, is challenging for scholars trained in social network analysis, a field that developed most of its techniques to describe small and sparsely featured datasets. As it is the case for their smaller counterparts, Big Relational Data is characterized by observations that are not independent and identically distributed draws from a population. Paraphrasing Tobler’s first law of geography, in Social Networks everything is related to everything else, but big connected things are more related to each other than big unconnected ones. In the case of BRD, data storing and representation, processing time, and model specification need to be addressed together if we are to successfully migrate from small and shallow relational datasets to large and fully featured ones.

In this paper, we introduce readers to a variety of strategies to model Big Relational Data. We propose a Path Weighted Local Regression (PWR) estimate to deal with the challenges of spatial dependence, structure, and network heterogeneity of big relational data. Inspired by the Geographically Weighted Regression Model, our PWR strategy describes the effect of unobserved factors across closely connected nodes. The model
estimates local parameters, with distinct estimates that characterize large networks with heterogeneous density regions.
Our model provides a computationally efficient strategy to model network structure, with estimates that grow linearly with network size and that can be easily parallelized. We exemplify the proposed method using Twitter data from the Travel Ban, the election of Judge Kavanaugh, and the election of Jair Bolsonaro.

Authors

• Ernesto F. Calvo, University of Maryland

• Joan C. Timoneda, Duke University

• Tiago Augusto da Silva Ventura, University of Maryland