Individual Submission Summary
Share...

Direct link:

Beyond Network Metrics: Graph Neural Networks for Sociological Network Analysis

Mon, August 10, 10:00 to 11:30am, TBA

Abstract

Social network analysis typically represents “structure” through an ever-expanding list of network metrics (e.g., degree, betweenness, density), often with the expectation that the right set of summaries will capture what matters about connectivity for shaping individual subjective experience. Yet across studies and contexts, relationships between these metrics and subjective outcomes are often unstable. Using friendship networks from 24,191 students in 56 New Jersey middle schools, we show that centrality predicts students’ sense of belonging in some schools but not others, and that observed school-level moderators (e.g., reciprocity) explain only part of this cross-school heterogeneity.

We ask whether this inconsistency reflects genuine context dependence or, instead, information loss introduced when connectivity is compressed into metrics. To address this question, we bring graph neural networks (GNNs) into sociological network analysis as a structure-preserving complement to metric-based models. We implement a stepped diagnostic that compares three approaches organized by how they handle connectivity: (1) multilevel logistic regression using demographic covariates, school context, and centrality measures; (2) gradient-boosted models using the same metric inputs to isolate gains from flexible functional form; and (3) a simple, strongly regularized GNN that operates directly on the adjacency structure plus node attributes, without centrality or school-level network summaries. Rather than treating predictive performance as a contest, we treat systematic divergence in predicted probabilities of belonging as evidence that the full adjacency structure contains incremental predictive information over standard metrics—and we identify the network contexts in which that information matters most.

Authors