How Influencers and Multipliers Drive Polarization and Issue Alignment on Twitter/X

TL;DR

The paper investigates polarization and issue alignment on the German Twittersphere (2021–2023) by combining BERTopic-based topic modeling with retweet-network clustering to infer user opinions and a novel alignment framework. It identifies two power-user roles—influencers (high in-degree) and multipliers (high out-degree)—that drive content generation and curated amplification, leading to strong cross-topic issue alignment and a highly polarized public sphere. The study shows multipliers exhibit stronger and more consistent issue alignment across topics than influencers, with certain topics like migration and Ukraine showing deviations, hinting at a nuanced cleavage beyond simple left-right dichotomy. These findings have implications for understanding online opinion formation and inform platform governance and policy discussions around content curation and amplification on social media.

Abstract

We investigate the polarization of the German Twittersphere by extracting the main issues discussed and the signaled opinions of users towards those issues based on (re)tweets concerning trending topics. The dataset covers daily trending topics from March 2021 to July 2023. At the opinion level, we show that the online public sphere is largely divided into two camps, one consisting mainly of left-leaning, and another of right-leaning accounts. Further we observe that political issues are strongly aligned, contrary to what one may expect from surveys. This alignment is driven by two cores of strongly active users: influencers, who generate ideologically charged content, and multipliers, who facilitate the spread of this content. The latter are specific to social media and play a crucial role as intermediaries on the platform by curating and amplifying very specific types of content that match their ideological position, resulting in the overall observation of a strongly polarized public sphere. These results contribute to a better understanding of the mechanisms that shape online public opinion, and have implications for the regulation of platforms.

Figures (10)

• Figure 1: Analysis pipeline. The raw text data of the tweets is processed in a topic model to extract the main issues discussed, which is then used to assign an overall issue to each trend. In parallel, retweet networks are constructed and clustered to extract the opinions of users for each trend. From this set of partitions, we extract the main actors (influencers and multipliers) and compute a user alignment matrix that quantifies how systematically pairs of users appear in the same cluster across the dataset. Importing the issue labels from the topic model analysis allows us to compute the user alignment separately for each issue, which we then use to generate the issue alignment matrix that measures how similarly two issues sort these users into opinion groups.
• Figure 2: Force-directed layout representation of two retweet networks. A shows a retweet network from a polarized trend, B one from an unpolarized trend.
• Figure 3: User alignment for influencers, multipliers and a random sample of users. We compute the pairwise user alignment for the top 1000 influencers (highest in-degree), top 1000 multipliers (highest out-degree) and a set of 1000 randomly sampled users. The matrices are sorted based on a hierarchical linkage clustering. If users do not participate in any same trend, the matrix field is left white. For each user set, there is a left and right-leaning cluster. In between, there are users that retweet and are retweeted by both opinion groups. Note that multipliers are more strongly aligned than the other groups and the right-leaning cluster is significantly larger. The matrix for the random sample shows that the division into the two camps exists here too, but the common participation in retweet networks is lower.
• Figure 4: Global and topic-wise cluster membership score for influencers (left) and multipliers (right). A membership score of $-1$ means that the user belongs to the left-leaning cluster (blue color), $+1$ to the right-leaning cluster (green color). Blank lines in the matrix mean that the user did not participate in any retweet network associated to the given topic. We observe that multipliers are more active across topics than influencers.
• Figure 5: Issue alignment for influencers (left) and multipliers (right). Both matrices are sorted according to optimal leaf ordering. For both user groups, we observe a strong issue alignment across topics, except for Music and Gaming. Multipliers exhibit a stronger issue alignment than influencers. There is no apparent cluster structure in either matrix.