Coordinated Behavior on Social Media in 2019 UK General Election

Contributions

• Non-binary coordination measurement: moves beyond existing threshold-based approaches by proposing an iterative framework that measures coordination as a continuous spectrum, enabling detection of weakly-coordinated communities that fixed thresholds would miss.
• Coordination-aware community detection: introduces an algorithm that performs community detection at increasing coordination levels, revealing how coordinated groups structure themselves across different extents of coordination.
• Comprehensive network characterization: applies multiple network measures (density, clustering coefficient, assortativity, degree distribution, centrality) to describe the structural properties of coordinated communities and how they vary with coordination extent.
• Large-scale public dataset: creates and publicly releases a Twitter dataset of 11.2 million tweets from 1.2 million users collected during the 2019 UK General Election period (November 12–December 12, 2019), including user interactions, hashtag data, and political leaning labels.
• Multi-faceted analysis: provides visual inspection, network measures, word shift analysis, and automation score analysis to characterize the themes, narratives, and composition (human vs. bot) of coordinated communities.

Method

Data collection: Leveraged Twitter Streaming APIs to collect tweets related to the 2019 UK General Election using a curated list of 13 political hashtags (e.g., #GE2019, #GeneralElection19, #VoteLabour, #ForTheMary, #ChangesComing). Dataset includes all tweets, retweets, and replies published during the 2-month pre-election window (November 12 – December 12, 2019), with 11.2M tweets shared by 1.2M users.

User similarity network:

1. User selection: Identified a set of seed users (superproducers and superspreaders) based on activity and reach.
2. Similarity measure: Computed pairwise user similarities using cosine similarity of user feature vectors, with two main approaches: - Hashtag similarity: measured shared hashtag usage - Interaction-based similarity: measured direct engagement patterns (retweets, mentions)
3. Network construction: Built a weighted undirected user similarity network encoding behavioral and interaction patterns between users.
4. Filtering: Applied multi-scale filtering to retain only statistically-relevant links, removing noise while preserving meaningful structures.

Coordination-aware community detection:

1. Iterative process: Performed community detection iteratively at increasing coordination thresholds, progressively filtering the network to retain only edges with similarity above each threshold.
2. Filtering algorithm: Used edge-weight thresholding to remove weak edges; for each threshold value, disconnected nodes were removed and communities re-detected.
3. Community detection: Applied standard algorithms (e.g., Louvain method) on the filtered networks to identify communities at each coordination level.
4. Coordination extent: Measured by the percentile rank of edge weights retained in the filtered network (e.g., coordination extent = 0.5 means retaining the top 50% of edge weights).

Network characterization: For each detected community, computed standard network measures:

• Density: fraction of possible edges present
• Clustering coefficient: transitivity of connections
• Assortativity: tendency of high-degree nodes to connect to other high-degree nodes
• Degree distribution
• Centrality measures (betweenness, closeness)

Results

Coordinated communities identified (Figure 4):

The analysis uncovered seven major coordinated communities, color-coded by political alignment:

1. CON: Conservative users (blue). Largest community with strong coordination (coordination extent ~0.75–0.9). Hashtags: #BackBoris, #VoteConservative, #ToriesInForTheLongGame
2. LAB: Labour users (red). Large, densely-coordinated community. Hashtags: #VoteLabour, #saveJeremy, #labourmanifesto, #ClinimateChange
3. TVT: Neutral/tactical voting users (orange). Large cluster related to liberal democrats and anti-Brexit campaigns. Hashtags: #VoteTactically, #RevokeArticle50, #stopbrexit, #stopcynumber
4. SNP: Scottish National Party supporters (yellow). Community focused on Scottish independence. Hashtags: #indyref2020, #SNPManifesto
5. B60: Small activist cluster (purple) involved in Back60 initiative (anti-austerity, women's rights movement). Strong coordination within small group.
6. ASE: Anti-establishment users (cyan) with strong Conservative-leaning sentiment.
7. LCH: Anti-semitism activists (brown) focused on attacks against Labour party leader.

Political leaning analysis (Figure 2): Computed political leaning scores for users based on hashtag polarity using a label propagation algorithm. The user network exhibits clear political polarization with distinct red (Labour), blue (Conservative), and orange (neutral/tactical) clusters.

Coordination patterns (Figure 7): Different communities exhibit strikingly different coordination profiles:

• High-coordination communities (LAB, CON): maintain significant coordination across broad ranges of thresholds (coordination extent 0.5–0.9)
• Weakly-coordinated communities (TVT, B60): show coordination only at lower thresholds (coordination extent < 0.3)
• Network measures vary non-monotonically: density typically increases then decreases with coordination extent; clustering coefficient and assortativity show distinct patterns for each community

Themes and narratives (Figure 8): Word shift analysis reveals distinct content production patterns:

• CON community: focused on party support, election turnout, Brexit
• LAB community: emphasized healthcare, antisemitism, inequality
• B60 and LCH communities: characterized by high automation and coordinated hashtag usage

Automation analysis (Figure 9): Correlation between coordination and automation was weak and community-dependent:

• Most coordinated users in some communities (CON, LAB) were human-operated
• Some communities (B60) showed higher bot proportions but this did not predict coordination extent
• Automation and coordination are orthogonal concepts rather than synonymous

Comparative evaluation: Compared findings against simpler threshold-based approaches (fixed edge-weight filtering). The binary classification approach (Figure 10) captures only a small portion of coordinated structure and entirely misses weakly-coordinated groups, demonstrating the advantage of the continuous measurement framework.

Notes

Strengths:

• Novel continuous coordination measurement: overcomes the limitations of binary coordinated/uncoordinated distinction by providing a nuanced, multi-scale view of coordination.
• Large-scale public dataset: 11.2M tweets with geographic and temporal coverage make this a valuable resource for future research on election-related coordination.
• Comprehensive multi-perspective analysis: combines network measures, thematic analysis, automation detection, and structural characterization to provide rich understanding of coordinated communities.
• Practical framework: the methodology is generalizable to other social media platforms, time periods, and events.
• Political polarization insights: clearly demonstrates how different political factions employ coordination strategies at different intensities and via different mechanisms.

Limitations:

• Platform and time-specific: results are specific to Twitter during a particular election period; generalization to other platforms or events may not be straightforward.
• Causality unclear: identifies coordinated patterns but does not establish whether coordination is spontaneous/organic (genuine shared interests) or orchestrated/inauthentic; distinguishing between authentic grassroots movements and coordinated campaigns remains challenging.
• Community definitions sensitive to parameters: thresholds and similarity measures influence which communities emerge; sensitivity analysis could be more thorough.
• Limited external validation: does not validate identified communities against known coordinated campaigns or ground truth labels (though this may be unavailable for recent events).