A Measure of Polarization on Social Media Networks Based on Community Boundaries

Pedro Guerra, Wagner Meira, Claire Cardie, Robert Kleinberg

Abstract

Polarization in social media networks is a pervasive phenomenon in political and social debates. Quantifying this polarization is critical for understanding online discourse and opinion analysis. While modularity is the standard metric for community detection, it often fails to distinguish between simple group segregation and active antagonism. We propose a novel polarization metric based on the analysis of community boundaries. Our findings indicate that in polarized networks, high-degree "popular" nodes avoid the boundaries, whereas in non-polarized networks, they act as bridges. We validate these metrics using datasets ranging from sports rivalries to the U.S. gun control debate.

Introduction: The Modularity Paradox

In network science, modularity ($Q$) measures the strength of division of a network into modules or clusters. A high $Q$ value suggests that nodes within a group are more densely connected to each other than to nodes in other groups. However, high modularity does not inherently imply polarization. For instance, fans of different sports teams may form distinct communities (high modularity) without being antagonistic toward one another. Conversely, political factions are often characterized by active "us vs. them" dynamics. To bridge this gap, we must look beyond the centers of communities and examine the "front lines" where they meet.

Methodology: The Boundary and the Polarization Metric

We define the Boundary ($B$) of a community as the subset of its nodes that possess at least one edge connecting to a different community. By focusing on these nodes, we can determine whether the interaction between groups is pluralistic or antagonistic.

Figure 4: Schematic representation of community boundaries between two groups.

We introduce the Polarization Metric ($P$), derived from the ratio of internal to external connections for boundary nodes. A positive $P$ value indicates that even nodes at the edge of a community prefer to link back to their own "echo chamber," signaling antagonism. A negative $P$ suggests a "pluralistic" overlap where nodes bridge the gap between communities.

Furthermore, we measure Boundary Concentration ($\rho$) using Spearman’s rank correlation. This metric assesses whether high-degree (popular) nodes are active on the boundary. In non-polarized networks, "stars" often serve as bridges ($\rho$ is high). In polarized contexts, these influential users remain deep within their own ideological territory, leaving the boundary to less popular nodes ($\rho$ is low).

Empirical Evidence: Antagonism vs. Pluralism

Our analysis across diverse datasets confirms that modularity is a deceptive proxy for polarization. For example, the Brazilian Soccer network (highly polarized) and the NYC Sports network (non-polarized) exhibited similar modularity scores, yet their polarization metrics were starkly different.

Figure 5: Distribution of node polarization for polarized (Soccer) and non-polarized (NY Sports) networks.

As shown in the comparison of $P_v$ distributions, the polarized soccer rivalry maintains a high positive polarization across many nodes, while the non-polarized sports network quickly drops into negative (pluralistic) territory.

In a case study of the Twitter gun control debate, our metric $P$ revealed nuanced dynamics that modularity missed. While three communities were identified, $P$ showed that while Conservatives were highly polarized against the other two groups, the Liberal and Independent groups actually shared a non-polarized, overlapping boundary ($P = -0.14$), despite being mathematically distinct clusters.

Conclusion

By shifting the focus to community boundaries, we provide a more accurate measure of social antagonism. Our metrics distinguish between benign segregation and harmful polarization, revealing that the "stars" of polarized debates are rarely the ones building bridges. This work offers a vital tool for researchers and platforms seeking to understand and mitigate the effects of echo chambers in digital society.

Summary updated

Abstract

Polarization in social media networks is a fact in several scenarios such as political debates and other contexts such as same-sex marriage, abortion and gun control. Understanding and quantifying polarization is a long-term challenge to researchers from several areas, also being a key information for tasks such as opinion analysis. In this paper, we perform a systematic comparison between social networks that arise from both polarized and non-polarized contexts. This comparison shows that the traditional polarization metric -modularity - is not a direct measure of antagonism between groups, since non-polarized networks may be also divided into fairly modular communities. To bridge this conceptual gap, we propose a novel polarization metric based on the analysis of the boundary of a pair of (potentially polarized) communities, which better captures the notions of antagonism and polarization. We then characterize polarized and non-polarized social networks according to the concentration of high-degree nodes in the boundary of communities, and found that polarized networks tend to exhibit low concentration of popular nodes along the boundary. To demonstrate the usefulness of our polarization measures, we analyze opinions expressed on Twitter on the gun control issue in the United States, and conclude that our novel metrics help making sense of opinions expressed on online media.
BibTeX
@inproceedings{guerra2021measure,
  author = {Pedro Guerra and Wagner Meira and Claire Cardie and Robert Kleinberg},
  title = {{A Measure of Polarization on Social Media Networks Based on Community Boundaries}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v7i1.14421},
}
Figure