Synchronization between media followers and political supporters during an election process: towards a real time study

TL;DR

This study analyzes how Twitter discussions around the 2022 French presidential election synchronized across political supporters and media-following groups. It introduces two dynamic semantic-network constructions—rolling window with bounded memory and growing aggregated networks—and measures inter-group similarities and entropy to detect synchronized attention in near real time. The work demonstrates that these methods reveal complementary insights, including short-lived bursts missed by static analyses, and provides a scalable approach for assessing equity in information treatment. Overall, it offers a practical framework for automatic, agnostic assessment of topic dynamics across diverse actor groups in political discourse.

Abstract

We present an analysis of the dynamics of discussions in Twitter (before it became X) among supporters of various candidates in the 2022 French presidential election, and followers of different types of media. Our study demonstrates that we can automatically detect the synchronization of interest among different groups around specific topics at particular times. We introduce two complementary methods for constructing dynamic semantic networks, each with its own advantages. The growing aggregated network helps identify the reactivation of past topics, while the rolling window network is more sensitive to emerging discussions that, despite their significance, may appear suddenly and have a short lifespan. These two approaches offer distinct perspectives on the discussion landscape. Rather than choosing between them, we advocate for using both, as their comparison provides valuable insights at a relatively low computational and storage cost. Our findings confirm and quantify, on a larger scale and in an automatic, agnostic manner, observations previously made using more qualitative methods. We believed this work represents a step forward in developing methodologies to assess equity in information treatment, an obligation imposed by law on broadcasters that use broadcast spectrum frequencies in certain countries.

Figures (12)

• Figure 1: Breakdown of the different types of followers for the official candidates (top) and selected media (bottom)."Followers": total number of accounts following the candidate/media on Twitter as of June 2022."Active": followers of the candidate/media who publish posts (tweets) or forward (retweet) those of other users. "in France": Followers with a self-declared location in France."Supporters": Twitter accounts where more than 75% of their retweets of candidates/media corresponding to a given candidate/media.
• Figure 2: Illustration of the updating scheme for the hashtags co-occurrence graph and the supporters’ hashtags usage
• Figure 3: Comparison of the evolution of the global network properties between the rolling window and growing aggregated co-occurrence graphs. The first four rows show standard network metrics: (a) number of nodes, (b) number of edges, (c) density, (d) clustering coefficient, (e) average degree, (f) degree assortativity, (g) number of connected components, (h) relative size of the largest component. The bottom row, shows in (i) the persistence of the co-occurrence graphs: fraction of nodes/edges present at the beginning of the capture, that are still present at time $t$ , and in (j) the "instantaneous" renewal of the graphs: fraction of nodes/edges present at time $t$ that were already present at time $t-1$.
• Figure 4: Hasthags' and Topics' entropy evolution. (a): Time evolution of the global hashtags' entropy. Notice that in this case the method of constructing the semantic network is irrelevant. (b): Comparison of the time evolution of the topics' entropy, computed using the rolling window and aggregated co-occurrence networks. The vertical lines mark the dates of occurrence of important specific events. (c): Comparative evolution of the number of topics between the two methods. For (b) and (c) the orange and blue curves correspond to the growing aggregated and rolling windows approaches, respectively.
• Figure 5: Time evolution of the self-similarity of the three candidates having obtained the largest amount of votes in the election Macron, Le Pen et Mélenchon along with that of the null model, for comparison. Vertical lines signal the begining of the Ukranian war and the date of the two rounds of the 2022, French presidential election. (a) Measurements corresponding to the rolling window network method. (b) Measurements corresponding to the aggregated growing network method. (c) Measurements corresponding to the static network method, for comparison