Finding polarised communities and tracking information diffusion on Twitter: The Irish Abortion Referendum

TL;DR

The paper tackles polarisation and information diffusion on Twitter during Ireland's Referendum by building a sentiment-weighted mentions network to identify two polarised groups with high accuracy. It demonstrates that a simplified approach using only the mentions network—avoiding followership data—achieves a balanced accuracy of $0.909$ in classifying Yes vs No supporters and enables robust cascade analysis. Using Goel et al.'s cascade reconstruction, the study shows information predominantly diffuses within the same ideological community, revealing a strong echo-chamber effect with minimal cross-community spread. The findings provide a scalable methodology for studying diffusion on large networks and offer insights into how polarised conversations propagate content, with implications for social-media analysis and policy-focused research. $M(T)$, $A(T)$, and $V(T)$ metrics corroborate the predominance of localized diffusion patterns across cascades.

Abstract

The analysis of social networks enables the understanding of social interactions, polarisation of ideas, and the spread of information and therefore plays an important role in society. We use Twitter data - as it is a popular venue for the expression of opinion and dissemination of information - to identify opposing sides of a debate and, importantly, to observe how information spreads between these groups in our current polarised climate. To achieve this, we collected over 688,000 Tweets from the Irish Abortion Referendum of 2018 to build a conversation network from users mentions with sentiment-based homophily. From this network, community detection methods allow us to isolate yes- or no-aligned supporters with high accuracy (90.9%). We supplement this by tracking how information cascades spread via over 31,000 retweet-cascades. We found that very little information spread between polarised communities. This provides a valuable methodology for extracting and studying information diffusion on large networks by isolating ideologically polarised groups and exploring the propagation of information within and between these groups.

Figures (17)

• Figure 2.1: Schematic of our methodological approach.
• Figure 3.1: Data summary for the Repeal the $8^\textrm{th}$ discussion. (a) Repeal the $8^\textrm{th}$ activity time frame showing the televised debates and the day of the referendum. Note a high peak on the referendum day. (b) Complementary cumulative distribution function for the number of tweets per user and an inset of the probability distribution function of the same data.
• Figure 3.2: (a) Complementary cumulative distribution function (CCDF) for in-degree (purple) and out-degree (green) per user in the mutual network, and inset of the probability distribution function (PDF) of the same network; (b) Average shortest path; (c) Local clustering coefficients.
• Figure 3.3: (a) Distribution of average sentiment-in by user in blue, and average sentiment-out by user in pink; (b) Distribution of simulated correlation values between the sentiment-in and sentiment-out by user, and the observed correlation. The blue bars are the resultant correlation distribution obtained after 1 000 simulations, and the red dashed line represents the observed correlation.
• Figure 3.4: Results of the randomisation test in the mutual mentions network. Green squares indicate that the observed fraction of connections falls outside the lower $2.5\%$ and upper $97.5\%$ quantiles of the randomised distribution (i.e., it is unlikely to arise by chance). We denote the fraction of links between negative and negative users as fnn, the fraction of links between negative and positive users as fnp, between positive and negative users as fpn, and between positive and positive users as fpp.