'Congratulations, morons': Dynamics of Toxicity and Interaction Polarization in the Covid Vaccination and Ukraine War Twitter Debates

TL;DR

This paper advances polarization research by treating polarization as a dynamic, multi-faceted process observable in real-time Twitter diffusion of two major public concerns: Covid vaccination and the Ukraine war. It combines time-resolved retweet diffusion, PCA/SVD-based diffusion spaces, density-based clustering, toxicity scoring via Perspective API, and Granger-causality analyses to reveal how influencer preferences and affective signaling evolve and interact. The findings show clear interactional polarization with temporally linked toxicity and structural divergence across cluster pairs, including notable cross-cluster dependencies and even polarization within ideological camps. These insights highlight the importance of dynamic, cross-variable analyses for understanding polarization's diffusion and its practical implications for information ecosystems and public discourse.

Abstract

The existence of polarization and echo chambers has been noted in social media discussions of public concern such as the Covid-19 pandemic, foreign election interference, and regional conflicts. However, measuring polarization and assessing the manner in which polarization contributes to partisan behavior is not always possible to evaluate with static network or affect measurements. To address this, we conduct an analysis of two large Twitter datasets collected around Covid-19 vaccination and the Ukraine war to investigate polarization in terms of the evolution in influencer preferences and toxicity of post contents. By reducing retweet behavior in each sample to several key dimensions, we identify clusters that reflect ideological preferences, along with geographic or linguistic separation for some cases. By tracking the central retweet tendency of these clusters over time, we observe differences in the relative position of ideologically unaligned clusters compared to aligned ones, which we interpret as reflecting polarization dynamics in the information diffusion space. We then measure the toxicity of posts and test if toxicity in one cluster can be temporally dependent on its structural closeness to (or toxicity of) another. We find evidence of ideological opposition among clusters of users in both samples, and a temporal association between toxicity and structural divergence for at least two ideologically opposed clusters in our samples. These observations support the importance of analyzing polarization as a multifaceted dynamic phenomenon where polarization dynamics may also manifest in unexpected ways such as within a single ideological camp.

Figures (6)

• Figure 1: Time series for Covid (left) and Ukraine (right) samples. Retweet counts are depicted with solid lines while the number of unique users is shown with the dashed curves. To account for differences in scale, each plot has separate y-axes for each variable.
• Figure 2: Time series plots for each pair of clusters in the Covid sample. Blue curves show the structural dissimilarity for those clusters as defined in equation \ref{['struct_sim']} while red curves show the combined toxicity for the two clusters as defined in equation \ref{['joint_toxicity']}. The shaded area hugging each curve corresponds to the difference between residuals from OLS fits and observations smoothed using a Gaussian kernel ($\sigma$ = 3). Each curve is plotted on its own y-axis with tick label colors corresponding to the respective curve colors. We include a light-blue horizontal dashed line to indicate the location of zero on the blue y-axis. A vertical black (solid) line indexes the beginning of the Ukraine war while a gray (dashed) vertical line shows where the time series is clipped for Granger tests.
• Figure 3: Time series plots for each pair of clusters in the Ukraine sample. Blue curves show the structural dissimilarity for those clusters as defined in equation \ref{['struct_sim']} while red curves show the combined toxicity for the two clusters as defined in equation \ref{['joint_toxicity']}. The shaded area hugging each curve corresponds to the difference between residuals from OLS fits and observations smoothed using a Gaussian kernel ($\sigma$ = 3). Each curve is plotted on its own y-axis with tick label colors corresponding to the respective curve colors.
• Figure 4: Left: Principal Component plots for the four retained components in the Covid sample. Color encodes users' cluster membership. Right: Bar plot showing cluster sizes. Colors in bar plot map to clusters in the PC plots.
• Figure 5: Hashtags with the top log-odds ratios per cluster (x-axis) for Covid (left) and Ukraine (right) samples. The log-odds ratio of a hashtag is computed in terms of the prevalence of that hashtag in a cluster compared to that hashtag's prevalence among users in all other clusters. Hashtag sets are ordered by cluster with the hashtags most idiosyncratic for C1 or U1 at the bottom and those for C4 or U5 at the top. Ratio markers are color encoded by cluster membership as specified in Figure \ref{['fig:cov_clust']} and Figure \ref{['fig:ukr_clust']}.