Polarization Detection on Social Networks: dual contrastive objectives for Self-supervision

Hang Cui, Tarek F. Abdelzaher

Introduction

Online social networks have become increasingly fragmented, characterized by the emergence of echo chambers and intense ideological polarization. Detecting these structures is critical for understanding social dynamics and mitigating the spread of misinformation. However, algorithmic detection remains a significant challenge. Most existing methods rely on "hand-crafted" features—such as specific toxic keywords or platform-specific interaction patterns (e.g., retweets)—which often fail to generalize across different platforms or topics. Furthermore, the scarcity of labeled data makes supervised learning difficult to scale.

To address these challenges, we present DocTra (Dual Contrastive objectives), a self-supervised framework designed to identify polarized groups without requiring manual labels or platform-specific feature engineering. DocTra leverages the inherent structure of social interactions and the "silence" between groups to learn robust representations of polarization.

Methodology: The DocTra Framework

DocTra operates on the principle that polarization is defined not just by who users interact with, but also by who they deliberately ignore. The framework employs two distinct contrastive learning objectives to capture these nuances.

1. Interaction-level Contrastive Objective: Modeling Silence

Traditional graph-based models focus on positive interactions (edges). DocTra introduces the concept of "polarization-induced silence." In a polarized environment, users often refrain from engaging with content that contradicts their worldview, even if that content aligns with their general topical interests.

The model identifies "hard negative" pairs: instances where a user would likely have engaged with a post based on general interest but chose not to. By contrasting actual interactions (positive pairs) with this deliberate silence (hard negative pairs), DocTra learns to map the boundaries of echo chambers more precisely than models that rely on random non-interactions.

2. Feature-level Contrastive Objective: Decoupling Bias from Interest

A significant hurdle in polarization detection is distinguishing between a shared interest (e.g., a local sports team) and a shared political stance. DocTra addresses this by decoupling node embeddings into two separate latent spaces:

  • Polarized Features: Capturing ideological bias and stance.
  • Invariant Features: Capturing general interests, locality, or topical relevance.

This decoupling ensures that the model's understanding of polarization is not "polluted" by non-polarized commonalities, allowing it to remain effective even when opposing groups discuss the same news events.

A Unified Polarization Index

Beyond detection, the framework introduces a Unified Polarization Index. Standard metrics often struggle to differentiate between a truly polarized social network and a naturally clustered network, such as a citation graph where researchers in the same field naturally cite one another. By scaling results using the "invariant features" identified during training, this new index provides a more accurate measure of social polarization relative to the underlying network structure.

Experimental Evaluation

The efficacy of DocTra was evaluated across seven diverse datasets, including:

  • Political Discourse: Twitter (TwPolitic), Reddit (RedditNews), and Chilean unrest data.
  • Social Issues: COVID-19 vaccination stances.
  • Collaborative Networks: Wikipedia editor communication (WikiTalk) and elections (WikiElec).

To ensure the model does not "hallucinate" polarization, it was also tested against five "unpolarized" benchmarks, such as the Cora and PubMed citation networks.

Key Results and Insights

DocTra demonstrated significant improvements over state-of-the-art (SOTA) baselines, including P-GNN and GraphMAE2. Key findings include:

  • Superior Accuracy: DocTra achieved the highest clustering accuracy across all seven datasets, showing a 5% to 10% improvement over previous self-supervised models.
  • Robustness to Noise: In semi-supervised settings, the model maintained stable performance even when 60% of the training labels were intentionally corrupted, highlighting the strength of its self-supervised objectives.
  • Signal in Silence: The results confirm that treating the absence of interaction as a deliberate signal is a powerful tool for mapping social boundaries.

Limitations and Future Directions

While DocTra represents a significant step forward, it faces certain limitations. The computational cost of modeling "silence" via link prediction can be high for massive graphs, though the authors suggest relaxation methods to mitigate this. Additionally, the current framework focuses primarily on binary polarization (e.g., a two-party system).

Future research will explore extending DocTra to multi-polar environments and investigating how these dual objectives can be used to identify "bridge-building" opportunities to reduce polarization in online communities.

Summary updated

Abstract

Echo chambers and online discourses have become prevalent social phenomena where communities engage in dramatic intra-group confirmations and inter-group hostility. Polarization detection is a rising research topic for detecting and identifying such polarized groups. Previous works on polarization detection primarily focus on hand-crafted features derived from dataset-specific characteristics and prior knowledge, which fail to generalize to other datasets. This paper proposes a unified self-supervised polarization detection framework, outperforming previous methods in unsupervised and semi-supervised polarization detection tasks on various publicly available datasets. Our framework utilizes a dual contrastive objective (DocTra): (1) interaction-level: to contrast between node interactions to extract critical features on interaction patterns, and (2) feature-level: to contrast extracted polarized and invariant features to encourage feature decoupling. Our experiments extensively evaluate our methods again 7 baselines on 7 public datasets, demonstrating significant performance improvements.
BibTeX
@inproceedings{cui2024polarization,
  author = {Hang Cui and Tarek F. Abdelzaher},
  title = {{Polarization Detection on Social Networks: dual contrastive objectives for Self-supervision}},
  booktitle = {CIC},
  year = {2024},
  doi = {10.1109/cic62241.2024.00020},
}
Figure