DISHONEST: Dissecting misInformation Spread using Homogeneous sOcial NEtworks and Semantic Topic classification

Caleb Stam, Emily Saldanha, Mahantesh Halappanavar, Anurag AcharyaarXiv 2024 · DOI · arXiv

Introduction

The "echo chamber" effect is a well-documented phenomenon in social media research, typically analyzed as a structural property of networks where like-minded individuals cluster together. However, this macro-level perspective often overlooks the nuances of individual user behavior. The DISHONEST paper addresses this gap by investigating the relationship between a user’s social interaction patterns—specifically, who they engage with—and their semantic diversity—what they talk about.

In the context of the COVID-19 "infodemic," understanding these dynamics is vital. If misinformation is propagated through isolated, semantically repetitive clusters, traditional mitigation strategies that focus solely on content removal may fail to disrupt the social mechanisms that sustain radicalization and community insulation.

Methodology

The researchers developed a multi-stage pipeline to integrate social network analysis with natural language processing (NLP).

  1. Graph Construction: A social network was modeled using retweets as edges, based on the premise that a retweet indicates agreement. Temporal snapshots were utilized to observe the network's evolution.
  2. Node Speed: The paper’s primary contribution is the "node speed" metric. It quantifies a user's propensity to interact with distant parts of the network. It is defined as the average shortest path length between a user and the new accounts they connect with during a specific time step. A "fast" user bridges disparate communities, while a "slow" user remains within their local clique.
  3. Topic Modeling: Using the Top2Vec algorithm, the authors identified 67 distinct topics within the tweet corpus. These were manually categorized into ten broader themes, such as "Conspiracy," "Policy," and "General anti-vax."
  4. Semantic Metrics: Two key metrics were established: "Topic Variation" (the breadth of topics a user discusses over time) and "Monotonicity" (the degree to which a user focuses on a single primary topic).
Figure 1: Top-level overview of the DISHONEST methodology pipeline, from dataset processing to inference.

The Avax Dataset

The study analyzed the Avax dataset, comprising over 1.2 million tweets related to vaccine hesitancy from October 2020 to May 2021. This period is particularly significant as it covers the lead-up to and the initial global rollout of COVID-19 vaccines. While the dataset includes over 568,000 unique users, the final analysis focused on a subset of 462 users with at least 50 retweets to ensure statistical robustness.

Key Insights and Results

The research reveals a significant social-semantic link: social diversity is a strong predictor of semantic diversity.

The Social-Semantic Correlation

Regression analysis between Node Speed and Topic Variation showed a highly significant relationship ($R^2 = 0.27, p < 0.001$). Users with higher "node speeds"—those reaching out to distant social groups—tend to discuss a wider variety of topics. Conversely, "slow" users often exhibit repetitive social behavior, such as retweeting a single account hundreds of times, and focus almost exclusively on a narrow set of narratives.

Isolation and Monotonicity

High "monotonicity" (repetitive focus on one topic) serves as a marker for social isolation. Approximately 70% of users who focused heavily on a single topic (monotonicity > 0.3) had abnormally low node speeds (below 1.5). This suggests that these users effectively isolate themselves even within broader misinformation subcommunities.

Network Evolution

Visual analysis of the social network shows a clear "densification" over time. As the pandemic progressed, the core of the misinformation network became increasingly interconnected and tightly packed.

Figure 2: Evolution of the largest connected component in the Avax social network across three timestamps (Dec 2020, Feb 2021, and May 2021), illustrating network densification.

Conclusion and Future Directions

The DISHONEST framework highlights that misinformation topics do not spread in isolation but rather in "bundles." Users engaging with one anti-vaccine narrative are likely to migrate through a cluster of related conspiratorial themes.

Future work will focus on refining the "node speed" metric across different platforms and mapping the influence of "fast" users to determine if they act as "super-spreaders" between clusters. Additionally, these insights could inform the development of "harm-aware" recommender systems designed to identify users in repetitive loops and introduce them to more diverse content.

Summary updated

BibTeX
@misc{stam2024dishonest,
  author = {Caleb Stam and Emily Saldanha and Mahantesh Halappanavar and Anurag Acharya},
  title = {{DISHONEST: Dissecting misInformation Spread using Homogeneous sOcial NEtworks and Semantic Topic classification}},
  year = {2024},
  eprint = {2412.09578},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.09578},
}
Figure