[Paper Review] Discovering Political Topics in Facebook Discussion threads with Graph Contextualization

Yilin Zhang, Marie Poux-Berthe+3 more|arXiv (Cornell University)|Aug 23, 2017

Complex Network Analysis Techniques28 references3 citations

TL;DR

This paper proposes pairGraphText, a spectral clustering method that integrates graph structure and high-dimensional text data to uncover political topics in Facebook discussion threads during the 2012 French presidential election. By contextualizing user interactions with textual content, the method successfully separates candidate-centered from issue-centered discussion structures, identifying key topics such as immigration, economy, and religion with high accuracy and scalability to large networks.

ABSTRACT

We propose a graph contextualization method, pairGraphText, to study political engagement on Facebook during the 2012 French presidential election. It is a spectral algorithm that contextualizes graph data with text data for online discussion thread. In particular, we examine the Facebook posts of the eight leading candidates and the comments beneath these posts. We find evidence of both (i) candidate-centered structure, where citizens primarily comment on the wall of one candidate and (ii) issue-centered structure (i.e. on political topics), where citizens' attention and expression is primarily directed towards a specific set of issues (e.g. economics, immigration, etc). To identify issue-centered structure, we develop pairGraphText, to analyze a network with high-dimensional features on the interactions (i.e. text). This technique scales to hundreds of thousands of nodes and thousands of unique words. In the Facebook data, spectral clustering without the contextualizing text information finds a mixture of (i) candidate and (ii) issue clusters. The contextualized information with text data helps to separate these two structures. We conclude by showing that the novel methodology is consistent under a statistical model.

Motivation & Objective

• To identify and distinguish between candidate-centered and issue-centered structures in political discussions on Facebook.
• To develop a scalable, unsupervised method that integrates high-dimensional text and network data for topic discovery in online political discourse.
• To improve clustering accuracy in noisy, high-dimensional text data by leveraging graph structure and text contextualization.
• To enable interpretable topic extraction from citizen comments using word-content strategies and interactive visualization tools.
• To provide a computationally efficient alternative to existing topic models like RTM for large-scale political network analysis.

Proposed method

• pairGraphText is a spectral clustering algorithm that jointly models user interaction networks (bi-partite between citizens and candidate posts) and textual content of comments.
• The method constructs a graph Laplacian from the interaction network and incorporates text features via a thresholded document-term matrix to reduce noise and dimensionality.
• It uses a joint optimization framework that enhances cluster homogeneity by aligning text and graph signals, improving signal detection in mixed structures.
• The algorithm applies spectral embedding to the combined graph-text matrix, followed by k-means clustering to identify latent discussion topics.
• It is designed to be computationally efficient and scalable, handling hundreds of thousands of nodes and thousands of unique words.
• The method is theoretically justified under a stochastic co-blockmodel, showing consistency in recovering true cluster labels under appropriate signal strength.

Experimental results

Research questions

• RQ1 How can we distinguish between candidate-centered and issue-centered discussion structures in Facebook political threads?
• RQ2 To what extent does integrating textual content improve the detection of political topics in networked discussion data?
• RQ3 Can a graph-based method effectively recover issue-centered topics when no ground-truth topic labels are available?
• RQ4 How does the performance of pairGraphText compare to relational topic models (RTM) and standard spectral clustering in mixed signal environments?
• RQ5 What are the most salient political topics attracting public attention in the 2012 French presidential election discussions on Facebook?

Key findings

• pairGraphText successfully separates candidate-centered and issue-centered structures in Facebook discussion threads, even when spectral clustering alone fails to disentangle them.
• In the absence of text signals, RTM fails to recover cluster labels, while pairGraphText maintains strong performance with graph signals alone, recovering over 90% of labels with sufficient signal strength.
• With both graph and text signals, pairGraphText achieves near-perfect cluster recovery, outperforming RTM in scenarios with high-dimensional, noisy text data.
• The method identifies key political topics such as immigration, economy, religion, ecology, and crises, which align with the salient issues of the 2012 French election.
• pairGraphText is significantly faster than RTM and more suitable for large-scale document clustering, while maintaining high accuracy and interpretability.
• The Shiny App and R package pairGraphText enable interdisciplinary collaboration and reproducible analysis, with full code and data available on GitHub.