Unpacking polarization: Antagonism and Alignment in Signed Networks of Online Interaction

TL;DR

The paper addresses how political polarization emerges in online signed interactions by separating hostility (Antagonism) from issue-structure alignment (Alignment). It introduces FAULTANA, a balance-theory-based pipeline that constructs signed networks, derives partitions with $k^*=2$, and computes $SAI_G$ and $SAI_R$ along with Cohesiveness and Divisiveness to quantify polarization. Applied to Birdwatch and DerStandard, the method reveals two dominant factions and identifies wedge issues, such as COVID-19 or refugees, that drive antagonism and alignment in platform-specific ways. The approach yields a time-resolved view of polarization dynamics and provides a transferable framework for analyzing polarity across platforms and languages.

Abstract

Political conflict is an essential element of democratic systems, but can also threaten their existence if it becomes too intense. This happens particularly when most political issues become aligned along the same major fault line, splitting society into two antagonistic camps. In the 20th century, major fault lines were formed by structural conflicts, like owners vs workers, center vs periphery, etc. But these classical cleavages have since lost their explanatory power. Instead of theorizing new cleavages, we present the FAULTANA (FAULT-line Alignment Network Analysis) pipeline, a computational method to uncover major fault lines in data of signed online interactions. Our method makes it possible to quantify the degree of antagonism prevalent in different online debates, as well as how aligned each debate is to the major fault line. This makes it possible to identify the wedge issues driving polarization, characterized by both intense antagonism and alignment. We apply our approach to large-scale data sets of Birdwatch, a US-based Twitter fact-checking community and the discussion forums of DerStandard, an Austrian online newspaper. We find that both online communities are divided into two large groups and that their separation follows political identities and topics. In addition, for DerStandard, we pinpoint issues that reinforce societal fault lines and thus drive polarization. We also identify issues that trigger online conflict without strictly aligning with those dividing lines (e.g. COVID-19). Our methods allow us to construct a time-resolved picture of affective polarization that shows the separate contributions of cohesiveness and divisiveness to the dynamics of alignment during contentious elections and events.

Figures (4)

• Figure 1: (Left) Schema of the FAULTANA pipeline: our analysis framework for Antagonism, Alignment, Cohesiveness, and Divisiveness. Grey boxes indicate data structures and variables implicated in the pipeline. Step 1 creates the relation network based on an aggregation of interactions through time. Step 2 applies the optimization algorithm, either exact or approximated, to obtain the optimal number of groups and optimal partition. From these two steps we can retrieve a global Alignment metric $SAI_G$. Then, by selecting subsets of the interaction data and with the optimal partition information, steps 3 and 4 compute the four metrics of interest: Antagonism, Alignment, Cohesiveness (normalized) and Divisiveness (normalized). (Right) Illustration of Antagonism and Alignment in signed network examples. The four networks have been constructed with the same edge density, number of nodes, and partitioning of nodes. Negative edges are red and positive edges are blue. The two upper networks have a higher proportion of negative edges, and thus higher Antagonism than the ones on the lower quadrants. Computed $SAI_{R(i)}$ values are provided to illustrate that the two right quadrants exhibit a higher level of Alignment, which is due to the lower amount of frustrated edges. Only the right upper quadrant corresponds to a strict definition of polarization in terms of both Antagonism and Alignment.
• Figure 2: (Upper figure) Signed network visualization of Birdwatch. Network of signed relationships for the BW1 dataset, comprising a total of 2,676 users and around 25,562 edges, negative colored red and positive blue. Node color corresponds to their group membership as identified by the exact method. Nodes belonging to the largest (smallest) group are depicted in yellow (black). Negative edges tend to connect different groups, while positive edges predominantly connect nodes within groups, demonstrating a considerable degree of balance. Insets: Inferred ideology of the targeted tweet's author separated by which group targeted the tweet and the nature of the note. We can only retrieve a score for tweet authors that have connections to political actors ($\sim 60\%$ of the users that posted tweets targeted in Birdwatch). The larger group gives misleading notes with more probability to tweets authored by Republican users, i.e. counter-partisan policing, with a slightly higher tendency to give not misleading notes to tweets by Democrat users. Thus, we identify the larger group as Democrat-leaning. The smaller group is much more likely to give not misleading notes to tweets authored by Republican users, showing a pattern of cheer-leading within Republicans and thus being identified as Republican-leaning. (Lower figure) Timeline of Alignment, Cohesiveness and Divisiveness in Birdwatch (BW1). The time series of each metric is calculated over a rolling window of ten days with increases of 5 days, with values allocated on the right of each window. The shaded area around Alignment time series shows 95% Confidence Intervals calculated against $10,000$ instances of the null model. Divisiveness is shown in red and Cohesiveness is shown in blue, with lighter areas showing the contribution of Democrat-leaning users to each metric and the remaining area above showing the contribution of Republican-leaning users. Bootstrapping intervals in Divisiveness and Cohesiveness are obtained for 10,000 bootstrap samples with replacement. The Alignment measure, $SAI_{R(t)}$, oscillates around a mean value of $0.65$. Divisiveness stays consistently above Cohesiveness, showing that negative interactions are the main driver of Alignment. Detected peaks in $SAI_{R(t)}$ are marked with circles and notable political events in the US are marked with vertical dashed lines for reference. For each peak, a summary text analysis of tweets in that period is shown in SI Appendix, Table S4, which can be further contextualized as increases in Cohesiveness, Divisiveness, or both. An interactive version of this plot can be found at https://emmafrax.github.io/BW1.html
• Figure 3: Alignment versus Antagonism and Cohesiveness versus Divisiveness across DerStandard topics. The left figure shows Antagonism and Alignment of the ratings of each news topic in DerStandard. Topics have been selected based on the topic/subtopic tags associated with the articles located above the postings (e.g., sports, climate change, etc.). Dashed lines show the mean values of each metric to identify the quadrants depicted in Figure \ref{['fig:pipeline']} An interactive version of this figure can be found at https://emmafrax.github.io/scatter.html. The right figure shows the scatterplot of Divisiveness versus Cohesiveness for DerStandard rating sub-sets based on topics. These two measures, which account for two different mechanisms that define Alignment, have a significant correlation across topics of 0.8. The highlighted outliers correspond to: (1) BVT (Austrian counterterrorism agency), (2) Abortion, (3) Scheuba (Austrian comedian) and (4) ÖVP (Political Party)
• Figure 4: Alignment timeline in DerStandard ratings sub-set of political topics, with detailed fluctuations in election periods. Upper timeline figure shows the Alignment measure obtained using a rolling window of 120 days of width and a step of 14 days. The features of the rolling window are selected so that the trends in Alignment through the eight years are visible, e.g. the change in trend at the start of 2016. In the lower figures we show more detailed changes of Alignment, with a rolling window of 30 days of width and a step of 7 days, around the three repetitions of the 2016 Presidential elections (A: 1,2 and 3) and the 2017 and 2019 Legislative Elections (B and C).