[Paper Review] Modeling Ideological Agenda Setting and Framing in Polarized Online Groups with Graph Neural Networks and Structured Sparsity.

Valentin Hofmann, Janet B. Pierrehumbert+1 more|arXiv (Cornell University)|Apr 18, 2021

Social Media and Politics120 references4 citations

TL;DR

This paper proposes a minimally supervised graph neural network with structured sparsity to detect ideological agenda setting and framing in polarized online communities using Reddit data. By modeling discourse structure and moral psychology insights, it identifies phenomena like radicalization and subreddit hijacking, achieving improved detection of polarized concepts across 12 years and 600+ online groups.

ABSTRACT

The increasing polarization of online political discourse calls for computational tools that are able to automatically detect and monitor ideological divides in social media. Here, we introduce a minimally supervised method that directly leverages the network structure of online discussion forums, specifically Reddit, to detect polarized concepts. We model polarization along the dimensions of agenda setting and framing, drawing upon insights from moral psychology. The architecture we propose combines graph neural networks with structured sparsity and results in representations for concepts and subreddits that capture phenomena such as ideological radicalization and subreddit hijacking. We also create a new dataset of political discourse covering 12 years and more than 600 online groups with different ideologies.

Motivation & Objective

• To develop a computational method for detecting ideological polarization in online discourse with minimal supervision.
• To model polarization along agenda setting and framing dimensions using insights from moral psychology.
• To create a large-scale, 12-year dataset of political discourse spanning 600+ online groups with diverse ideologies.
• To identify structural phenomena such as ideological radicalization and subreddit hijacking through learned representations.

Proposed method

• The method employs graph neural networks (GNNs) to model the network structure of Reddit discussion forums, capturing interactions between users and subreddits.
• Structured sparsity is applied to the GNN's attention mechanisms to identify salient concepts and subreddits associated with ideological polarization.
• The model jointly learns representations for both concepts and subreddits, enabling detection of how topics are framed and prioritized across ideological lines.
• It leverages moral psychology frameworks to guide the identification of polarized framing and agenda-setting behaviors in discourse.
• The architecture is trained in a minimally supervised manner, using only weak supervision signals from known ideological groupings.
• The method produces interpretable, low-dimensional representations that reflect ideological divergence and structural shifts in online communities.

Experimental results

Research questions

• RQ1 How can graph neural networks with structured sparsity detect polarized agenda setting in online political discourse?
• RQ2 To what extent can moral psychology-informed framing patterns be captured through representation learning in social media networks?
• RQ3 How do ideological radicalization and subreddit hijacking emerge in the learned representations of the model?
• RQ4 Can the model identify and distinguish between agenda-setting and framing behaviors across different ideological subreddits?
• RQ5 What is the performance of the method in detecting polarization compared to existing unsupervised or weakly supervised approaches?

Key findings

• The model successfully identifies polarized concepts and subreddits associated with agenda setting and framing, demonstrating strong alignment with known ideological divides.
• Structured sparsity enables the model to highlight key ideological concepts and subreddits, improving interpretability and reducing noise in representation learning.
• The method detects instances of ideological radicalization and subreddit hijacking through shifts in learned representations over time.
• The model outperforms baseline methods in detecting polarization with minimal supervision, as shown by qualitative and quantitative analysis on the new 12-year dataset.
• The learned representations capture nuanced differences in how topics are discussed across ideological lines, reflecting distinct framing strategies.