Dynamics of Ideological Biases of Social Media Users

TL;DR

This paper investigates how online social platforms shape ideological biases through homophily-driven dynamics, using Parler and Twitter data around the 2020 U.S. election. It develops a bias-mapping framework with eight categories and two constellations, and employs a Flow Matrix to track initial-to-final bias transitions while accounting for dropouts. The results reveal Twitter’s bimodal polarization with two stable bias groups and Parler’s dominant Right/Fake News echo chamber, highlighting platform-specific polarization patterns and movement dynamics. These findings offer design insights for socially aware networks and caching strategies, and motivate future graph-based and agent-based modeling to scale bias dynamics across platforms.

Abstract

Humanity for centuries has perfected skills of interpersonal interactions and evolved patterns that enable people to detect lies and deceiving behavior of others in face-to-face settings. Unprecedented growth of people's access to mobile phones and social media raises an important question: How does this new technology influence people's interactions and support the use of traditional patterns? In this article, we answer this question for homophily-driven patterns in social media. In our previous studies, we found that, on a university campus, changes in student opinions were driven by the desire to hold popular opinions. Here, we demonstrate that the evolution of online platform-wide opinion groups is driven by the same desire. We focus on two social media: Twitter and Parler, on which we tracked the political biases of their users. On Parler, an initially stable group of Right-biased users evolved into a permanent Right-leaning echo chamber dominating weaker, transient groups of members with opposing political biases. In contrast, on Twitter, the initial presence of two large opposing bias groups led to the evolution of a bimodal bias distribution, with a high degree of polarization. We capture the movement of users from the initial to final bias groups during the tracking period. We also show that user choices are influenced by side-effects of homophily. Users entering the platform attempt to find a sufficiently large group whose members hold political biases within the range sufficiently close to their own. If successful, they stabilize their biases and become permanent members of the group. Otherwise, they leave the platform. We believe that the dynamics of users' behavior uncovered in this article create a foundation for technical solutions supporting social groups on social media and socially aware networks.

Figures (3)

• Figure 1: A two-level clustering of polarized users. The lower level contains eight bias groups. The higher level consists of two primary clusters called constellations that group associated biases together. The Center bias and Fake news groups exist outside the two constellations. Edges connecting constellation's groups show that members can directly reach groups within each constellation, defining unit distance between them. Travel between groups across the constellations requires several unit steps. Each user has two biases, initial and final. The initial bias uses URL links from the initial month of collected data, while the final bias uses links gathered in the last month. Each user with two different biases travels from initial to final bias, changing the sizes of bias groups dynamically.
• Figure 2: Flow diagram of Twitter (Top) and Parler (Bottom) users. Column I shows the number of newcomers in each of the initial bias groups. Column D to the left of I shows the number of newcomers that drop out from the platform. Column A shows the number of newcomers who obtain a final bias classification. The Flow Matrix FM connects active users with the same initial bias to the final bias assigned to them. The bottom row F shows the number of users with their final biases. Thus, the direction of flow is from column I to A, then to columns FM along the corresponding row, and finally to row F.
• Figure 3: Diagram visualizing the movements for each bias group on Twitter and Parler. Each box plot shows the Interquartile Range for the initial-to-final bias group distances traveled by each user initially at that group. The yellow lines in the box plots represent the median distance traveled by the group members and whiskers on either side visualize the maximum extent of distance moved.