PaperlayerView exact revision on arXiv

Source-linked AI summary

Coordinated Behavior on Social Media in 2019 UK General Election

Leonardo Nizzoli, Serena Tardelli, Marco Avvenuti, Stefano Cresci, Maurizio Tesconi

arXiv:2008.08370v2cs.SI

TL;DR

Existing computational studies often reduce coordination to a binary classification, despite its complexity. This paper proposes a network framework that detects coordination patterns and estimates coordination across communities. Applied to the 2019 UK General Election, it found coordination in every online community, with the most coordinated groups being small activist and political-antagonist communities.

  • Problem

    Existing computational studies often use fixed thresholds to classify coordination as binary, although coordination is a complex, non-binary concept.

    Source: Introduction · p. 2

  • Method

    The framework filters a user-similarity network to retain multiscale structures and uses coordination-aware community detection to estimate coordination across communities.

    Source: Method overview · p. 4

  • Results

    The analysis found coordination in every online community; Conservatives were overall more coordinated than Labour, while the most coordinated groups were small activist and political-antagonist communities.

    Source: Conclusions · p. 11

  • Takeaways & Limitations

    Estimating coordination by community opens up more accurate and fine-grained downstream analyses than applying a single fixed threshold to the whole network.

    Source: Conclusions · p. 11

    A single threshold may be unsuitable across similarity representations, since a value suitable for hashtag similarity can be too restrictive for co-retweet similarity.

    Source: Comparative evaluation · p. 11

Abstract

from arXiv · show

Coordinated online behaviors are an essential part of information and influence operations, as they allow a more effective disinformation's spread. Most studies on coordinated behaviors involved manual investigations, and the few existing computational approaches make bold assumptions or oversimplify the problem to make it tractable. Here, we propose a new network-based framework for uncovering and studying coordinated behaviors on social media. Our research extends existing systems and goes beyond limiting binary classifications of coordinated and uncoordinated behaviors. It allows to expose different coordination patterns and to estimate the degree of coordination that characterizes diverse communities. We apply our framework to a dataset collected during the 2019 UK General Election, detecting and characterizing coordinated communities that participated in the electoral debate. Our work conveys both theoretical and practical implications and provides more nuanced and fine-grained results for studying online information manipulation.

Introduction

Information operations threaten democratic societies, and their reach often depends on coordinated social media efforts. Because coordination is distinct from authenticity and difficult to define, existing computational approaches’ binary assumptions leave important challenges unresolved.

Introduction

  • Information operations use deceptive content to mislead and manipulate, and their threat to democracies and societies persists despite extensive research and proposed solutions.They can involve fake news, propaganda, hateful speech, astroturfing, colluding users, or automation.Source: Introduction · p. 1
  • Campaigns often require extensive coordination to spread widely, but coordination and inauthenticity are distinct: grassroots activism can be coordinated yet authentic, while a misleading fake account can be inauthentic yet uncoordinated.The passage also notes that deepfakes have been used to create fake news and fraudulent profile pictures.Source: Introduction · p. 1
  • Computational study of coordination remains limited by its ambiguous definitions and by approaches that impose fixed thresholds and binary labels, despite coordination being non-binary.The authors focus on coordinated behavior, moving beyond binary approaches; automatic distinction between authentic and inauthentic coordination has not been reported.Source: Introduction · pp. 1–2

Contributions

The paper proposes a network-based framework that measures coordination continuously rather than classifying users as coordinated or not, and applies it to Twitter data from the 2019 UK General Election. It uses the framework to characterize coordinated communities and compare its approach with existing binary methods.

Contributions

  • The framework measures coordination on a continuous, non-binary scale without fixed thresholds, using a user-similarity network, a multi-scale backbone, and iterative community detection.It characterizes users from weakly to strongly coordinated and crosschecks its coordination index against network measures.Source: Contributions · p. 2
  • Applied to Twitter data from the 2019 UK General Election, the framework uncovers coordinated communities and discusses their role in the electoral debate.Source: Contributions · p. 2
  • The analysis identifies different coordination patterns across communities, enabled by the framework’s non-binary approach.Source: Contributions · p. 2
  • The paper empirically finds that coordination and automation are orthogonal concepts, so the framework can complement techniques for detecting automation, manipulation, and inauthenticity.Source: Contributions · p. 2
  • The authors compare the framework with existing binary approaches, reporting advantages in quantifying coordination extent and uncovering coordinated communities’ network properties.Source: Contributions · p. 2
  • The paper creates and publicly shares a 2019 UK General Election dataset containing 11M tweets from 1.2M users.Source: Contributions · p. 2

Related Work

Earlier computational studies often used fixed thresholds to identify coordinated groups, while this work examines coordination continuously. Related evaluations also show that coordination and inauthenticity are distinct concepts.

  • Unlike Pacheco et al. (2020; Pacheco, Flammini, and Menczer 2020), who threshold-filtered a user-similarity network into coordinated components, this work detects communities across coordination levels.This continuous measure reveals coordination patterns and dynamics that threshold-based approaches may miss, and can be crosschecked against network properties.Source: Related Work · p. 2
  • Giglietto et al. (2020b; 2020a) detected coordinated link-sharing groups using near-simultaneity and repetition thresholds, then manually assessed inauthenticity through shared domains and stories.The two fixed similarity thresholds introduce the limitations associated with threshold-based approaches.Source: Related Work · pp. 2–3
  • Assenmacher et al. (2020b; 2020a) used unsupervised stream clustering and trend detection to find similar users, followed by offline analyses, including manual inspection, to assess inauthenticity.Source: Related Work · p. 3
  • Keller et al. (2020) found coordination traces more informative than individual account traits and developed astroturfing detection, while Fazil and Abulaish (2020) labeled similarity-graph clusters as inauthentic coordinated groups.Fazil and Abulaish treated high coordination as implying inauthenticity.Source: Related Work · p. 3
  • Vargas, Emami, and Traynor (2020) found that prior network-based coordination metrics could not distinguish authentic from inauthentic coordination in non-trivial real-world scenarios.Their results indicate that high coordination does not necessarily imply inauthenticity.Source: Related Work · p. 3

Dataset

The study combines hashtag-based and official-account Twitter data collected during the month before the 2019 UK General Election. The resulting dataset contains 11,264,820 tweets by 1,179,659 distinct users and is publicly available for research.

  • From 12 November to 12 December 2019, researchers collected tweets mentioning popular election-related hashtags and official party and leader account posts, plus their retweets and replies.Source: Dataset · p. 3
  • The combined dataset contains 11,264,820 tweets by 1,179,659 distinct users and is publicly available for research.Source: Dataset · p. 3

Method overview

The framework builds a user-similarity network, filters it to retain meaningful structures across scales, and iteratively detects communities across increasing coordination levels. This makes it possible to characterize communities as a function of their measured coordination.

  • Multiscale network filtering retains statistically meaningful structures across edge-weight scales instead of imposing a fixed threshold that divides users into coordinated and uncoordinated groups.This avoids bias toward particular similarity levels and removes structures that convey limited information.Source: Method overview · p. 4
  • The first three steps select users and a similarity measure, then compute pairwise similarities to build a weighted, undirected user-similarity network.Similarity options include account-profile features, hashtags, followings, and retweeted accounts.Source: Method overview · pp. 3–4
  • Iterative community detection tracks how community structure and properties change across increasing coordination levels, using each iteration’s communities to initialize the next.The increasing similarity threshold removes edges and disconnected nodes; it also provides a measure of coordination at each iteration.Source: Method overview · p. 4
  • The final step studies community structure and coordination patterns with network measures, contextualizes communities, and analyzes content production using natural language processing.These analyses can be reported as a function of the measured extent of coordination between users.Source: Method overview · p. 4

Surfacing coordination in 2019 UK GE

The filtered user-similarity network revealed politically distinct communities in the 2019 UK General Election debate, including major party groups and smaller organized campaigns. Community detection exposed groups across differing levels of coordination, while the extent and patterns of coordination were left for the next section to examine.

User similarity network

  • 10,782 superspreaders—defined as the top 1% of users sharing the most retweets—shared 39% of all tweets and 44.2% of retweets.Users were represented by TF-IDF-weighted vectors of retweeted tweet IDs, with cosine similarity used to build the network; the filtered network contained 276,775 edges.Source: Surfacing coordination in 2019 UK GE · p. 5

Coordinated communities

  • Multiscale backbone filtering retained statistically relevant edges and the network’s rich, multiscale structure rather than applying a fixed edge-weight threshold.The resulting network included different degrees of similarity and coordination; Louvain communities were detected at resolution 1.5 with minimum size 20 at t0.Source: Surfacing coordination in 2019 UK GE · pp. 5–6

Network interpretation

  • The network separated into politically homogeneous communities, with Conservatives sharply apart from the more intertwined Labour and Neutral groups.The authors relate this structure to Brexit polarization and tactical voting in the UK electoral context.Source: Surfacing coordination in 2019 UK GE · pp. 5–6

Coordinated communities

  • Louvain community detection identified CON, LAB, TVT, and SNP groups, whose hashtags reflected party support, anti-Tory or anti-Brexit debate, and tactical voting.CON hashtags supported the Conservatives, Boris Johnson, and Brexit; LAB hashtags supported Labour, its leader, the NHS, and climate action, with no Brexit-related keywords. TVT covered Liberal Democrats, anti-Tory and anti-Brexit topics, and tactical voting; SNP users discussed the party and independence.Source: Surfacing coordination in 2019 UK GE · p. 6
  • The small B60 community comprised Backto60 activists and was linked to both LAB and TVT, reflecting the activists’ connections in the network.The campaign concerned women born in the 1950s who were affected by state pension age equalisation.Source: Surfacing coordination in 2019 UK GE · pp. 6–7
  • ASE users leaned Conservative but focused on attacking Labour and Jeremy Corbyn, especially over antisemitism allegations.Its prominent hashtags included labourlies, nevercorbyn, labourantisemitism, and votelabourvoteracism.Source: Surfacing coordination in 2019 UK GE · p. 7
  • The LCH cluster comprised activists protesting the loan charge, and the analysis found smaller clusters associated with organized activists and political campaigns.The authors characterize B60 and LCH as activist clusters and ASE as a political campaign.Source: Surfacing coordination in 2019 UK GE · p. 7
  • The detected communities showed at least some coordination, but the network analysis did not yet establish its extent or the coordination patterns characterizing those groups.The authors reserve analysis of coordination degree and patterns for the next section.Source: Surfacing coordination in 2019 UK GE · p. 8

Analysis of coordinated behaviors

The analysis goes beyond detecting coordinated communities to estimate coordination extent and investigate distinct network, content, and automation patterns within them.

Network measures

  • Unlike detection-only pipelines, the framework estimates coordination extent and investigates coordination patterns for each uncovered community.It repeatedly removes edges from weak to strong; a coordination value of 0.9 denotes a subnetwork containing only the top 10% strongest edges.Source: Analysis of coordinated behaviors · p. 8

Visual inspection

  • Visual inspection shows LAB and CON contain many weakly coordinated users alongside a smaller strongly coordinated core, while activist communities appear mostly strongly coordinated.SNP and TVT appear mildly coordinated, with only a few dispersed strongly coordinated users; B60, LCH, and ASE form small, compact dark clusters.Source: Analysis of coordinated behaviors · p. 8

Network measures

  • LCH and ASE retain members at high coordination levels, while B60, LAB, and TVT decline across the spectrum, indicating distinct characteristic coordination values.The elbow method suggests values of approximately 0.9 for LCH and 0.55 for ASE; applying one fixed threshold to every community risks erasing relevant patterns.Source: Analysis of coordinated behaviors · p. 8
  • Density and clustering vary by community: SNP density falls with coordination, B60 density spikes at ≥0.8, and clustering falls in SNP/LCH/TVT but rises in CON/LAB/B60.Together, density and clustering trends support a representative coordination value of approximately 0.9 for LCH.Source: Analysis of coordinated behaviors · pp. 8–9, Figure 7 · p. 9
  • ASE and LAB are moderately disassortative, consistent with hub-and-spoke structures, whereas B60 is strongly assortative near 0.8, consistent with coordinated peers connecting to one another.Combining the network measures, the authors describe B60 by a coordination value of approximately 0.8.Source: Analysis of coordinated behaviors · p. 9, Figure 7 · p. 9

Use of automation

  • Botometer scores and Twitter suspensions are largely uncorrelated with coordination, although LCH has higher automation scores, CON and ASE more suspensions, and B60 suspensions decline.The authors interpret this as evidence that coordination and automation are distinct, and that bot detection alone may be ineffective for studying coordinated inauthentic behavior.Source: Figure 9 · p. 10, Analysis of coordinated behaviors · p. 10

Themes and narratives

  • Word-shift analysis finds that coordinated users embrace more specific narratives than other community members, including 1950s women’s rights and opposition to the loan charge tax.The comparison uses all community tweets as a reference and tweets from strongly coordinated users as the comparison corpus; B60 and LCH thresholds are approximately 0.8 and 0.9, respectively.Source: Analysis of coordinated behaviors · pp. 9–10, Figure 8 · p. 9

Comparative evaluation

The comparison shows that fixed edge-weight thresholds can discard information and produce uneven group boundaries, while the framework’s multiscale, coordination-aware analysis preserves network structure and characterizes coordination across a spectrum.

Comparative evaluation

  • By avoiding a binary coordinated-versus-uncoordinated split, the framework measured each emerging community’s intrinsic coordination and characterized it across multiple dimensions.The paper contrasts this spectrum-wide analysis with approaches focused only on the most coordinated accounts.Source: Comparative evaluation · pp. 10–11
  • At threshold 0.9, only six edges survived, whereas the multiscale backbone retained significant network structures while keeping information manageable.The strict threshold ensured surviving edges were undoubtedly related to coordinated behaviors, but discarded much information.Source: Comparative evaluation · p. 10, Figure 10 · p. 11
  • The coordination-aware algorithm surfaced coordinated groups within communities, enabling analysis of shifts in supported themes and narratives and their interactions with other users.The continuous coordination measure also allowed groups to be characterized with informative network measures at varying coordination levels.Source: Comparative evaluation · pp. 10–11
  • Threshold suitability depends on the similarity representation, and relaxing it from 0.9 to 0.7 or 0.5 enlarged groups unevenly, especially the CON cluster relative to B60 and LAB.A threshold appropriate for hashtag-based similarity may be too restrictive for the less dense co-retweet representation.Source: Comparative evaluation · p. 11, Figure 10 · p. 11

Conclusions

The paper presents a network-based framework for analyzing the spectrum of coordinated behavior and applies it to the 2019 UK General Election. It finds coordination across communities but notes that the most coordinated groups were small activist and political-antagonist groups, while the framework still cannot distinguish authentic from inauthentic coordination.

Conclusions

  • The authors created and shared a Twitter dataset for the 2019 UK General Election and applied their framework to study coordinated behavior in that election.Source: Conclusions · p. 11
  • The framework moves beyond binary coordination labels by combining improved network filtering with coordination-aware community detection to reveal varied patterns across the coordination spectrum.It estimates coordination separately for each community rather than applying one fixed threshold across the network, enabling finer-grained downstream analyses.Source: Conclusions · p. 11
  • The analysis found some degree of coordination in every online community, but not all coordinated groups were equally interesting.Source: Conclusions · p. 11
  • Conservatives were overall more coordinated than Labour, with higher automation and Twitter suspensions, while the most coordinated communities were small activist and political-antagonist groups.These findings came despite smaller numbers of Conservative accounts.Source: Conclusions · p. 11
  • The framework does not yet distinguish inauthentic coordinated behavior from authentic coordination, although the authors describe its more nuanced results as progress toward that goal.Source: Conclusions · p. 11

Related papers

Loading 2008.08370v2…