Susceptibility of Communities against Low-Credibility Content in Social News Websites

TL;DR

This work develops a method to identify communities within a social news website that are prone to uncredible or highly biased news, using a user embedding pipeline that detects user communities based on their stances towards posts and news sources.

Abstract

Social news websites, such as Reddit, have evolved into prominent platforms for sharing and discussing news. A key issue on social news websites sites is the formation of echo chambers, which often lead to the spread of highly biased or uncredible news. We develop a method to identify communities within a social news website that are prone to uncredible or highly biased news. We employ a user embedding pipeline that detects user communities based on their stances towards posts and news sources. We then project each community onto a credibility-bias space and analyze the distributional characteristics of each projected community to identify those that have a high risk of adopting beliefs with low credibility or high bias. This approach also enables the prediction of individual users' susceptibility to low credibility content, based on their community affiliation. Our experiments show that latent space clusters effectively indicate the credibility and bias levels of their users, with significant differences observed across clusters -- a $34\%$ difference in the users' susceptibility to low-credibility content and a $8.3\%$ difference in the users' susceptibility to high political bias.

Figures (3)

• Figure 1: Overview of methodology: This diagram illustrates our analysis approach. Rounded boxes depict various processes, while sharp-edged boxes depict datasets. Diagonally upward arrows behind processes indicate the use of corresponding datasets for training these processes. The pooling blocks, indicated with a $\boldsymbol{+}$ sign denote averaging the inputs for each user.
• Figure 2: Alignment cost across different numbers of clusters for users' latent space embeddings.
• Figure 3: User distributions of all $13$ clusters across years 2016-2018. A: Latent space embedding visualization of users using UMAP reduction. B: Credibility-bias mappings of all users. Larger numbers denote higher credibility and right-wing political bias in their respective axes. C: Credibility-bias distributions of $9$ clusters with least maximal covariance eigenvalue, together with marginal distributions.