Held-Out Predictive Evaluation (HOPE) of Exponential Family Random Graph Models

• In Event: 4258 - Social Network Analysis: New Methodologies

Abstract

Model assessment for social network analyses are inherently difficult due to the relational nature of the data. As a result, statistical tools used for assessing models are not directly applicable. Yet, there remains the need for network scholars to assess how well their models are able to reproduce their original data. This paper fills that gap by developing a within-sample validation method for exponential family random graph models that is similar to leave-one-out cross-validation. Our algorithm systematically holds out parts of the network to evaluate: how well the model is able to estimate the held out data; where the model performs poorly based on held-out data, which indicates an idiosyncratic edge formation process for a node; and a graph level summary of the accuracy of the imputations. Our algorithm is a held-out predictive evaluation or HOPE. HOPE can assist researchers in improving models by indicating where a model performs poorly, and it facilitates the comparison of fit across multiple models. As such, the HOPE routine provides a valuable tool for modeling social networks.

Authors

• Nolan Phillips, Harvard University

• Carter T. Butts, University of California-Irvine