SGP: Spotting Groups Polluting the Online Political Discourse

Junhao Wang, Sacha Levy, Ren Wang, Aayushi Kulshrestha, Reihaneh Rabbany

Abstract

Recent events have led to a burgeoning awareness on the misuse of social media sites to affect political events, sway public opinion, and confuse the voters. Such serious, hostile mass manipulation has motivated a large body of works on bots/troll detection and fake news detection, which mostly focus on classifying at the user level based on the content generated by the users. In this study, we jointly analyze the connections among the users, as well as the content generated by them to Spot Coordinated Groups (SCG), sets of users that are likely to be organized towards impacting the general discourse. Given their tiny size (relative to the whole data), detecting these groups is computationally hard. Our proposed method detects these tiny-clusters effectively and efficiently. We deploy our SCG method to summarize and explain the coordinated groups on Twitter around the 2019 Canadian Federal Elections, by analyzing over 60 thousand user accounts with 3.4 million followership connections, and 1.3 million unique hashtags in the content of their tweets. The users in the detected coordinated groups are over 4x more likely to get suspended, whereas the hashtags which characterize their creed are linked to misinformation campaigns.

Introduction

The digital "town square" is increasingly populated by coordinated actors—trolls, sock puppets, and botnets—designed to sway elections and polarize public opinion. While individual bot detection has improved, identifying coordinated groups that act in synchronization remains a significant challenge. Most existing research focuses on user-level or content-level detection. However, modern information operations involve groups that may appear human-like individually but exhibit suspicious, synchronized behavior as a collective.

The technical challenge lies in scale and resolution. These coordinated groups are often "tiny clusters," representing less than 1% of the total dataset. Standard community detection algorithms often suffer from a "resolution limit," merging these small, high-density groups into larger, unrelated clusters, effectively rendering the manipulators invisible.

The SCG Framework

The SCG (Spotting Coordinated Groups) framework is a modular system designed to find these "needles in the haystack" by fusing social network structure with content analysis. It consists of four components:

  1. SCG-map: A fusion engine using truncated SVD and graph message passing to create joint embeddings of users based on followers (connections) and hashtags (content).
  2. SCG-flag: A density-based clustering module (HDBSCAN) that identifies tiny clusters that are both densely connected and messaging-similar.
  3. SCG-signature: This identifies the "group creed" by calculating hashtags uniquely popular within a cluster compared to the general population.
  4. SCG-meso: An analysis tool for "meso-level" interactions between different coordinated groups and the broader discourse.

Performance and Scalability

SCG was evaluated against several baselines using large-scale synthetic data. As shown in the performance table below, SCG maintains exceptionally high F1 scores even as the dataset size ($n$) increases, significantly outperforming traditional methods like Louvain or node2vec, which often fail to scale or lose precision in detecting tiny clusters.

Table 1: Performance comparison (Quality and F1) of SCG against baselines across varying dataset sizes.

Notably, SCG scales linearly with the number of edges. In empirical tests, it proved to be over 10,000 times faster than node2vec on medium-sized graphs, making it viable for real-time monitoring of massive social networks.

Case Study: 2019 Canadian Federal Election

The framework was deployed on a real-world dataset of ~70,000 Twitter accounts during the 2019 Canadian Federal Election. SCG successfully identified several distinct coordinated groups characterized by specific political agendas.

Figure 1: Visualization of detected coordinated groups and their associated word clouds (creeds) from the 2019 Canadian Federal Election data.

The analysis revealed striking interactions, such as American conservative groups (e.g., #WWG1WGA) interacting with Canadian anti-Trudeau groups. This suggests that "foreign interference" can manifest as horizontal interactions between ideologically aligned groups across borders rather than just top-down state operations.

Validation of these findings was stark: users in the groups flagged by SCG were four times more likely to be suspended by Twitter than the average user. Furthermore, SCG identified hashtags linked to misinformation campaigns (like #notAbot) before they were widely reported in mainstream media.

Conclusion

SCG demonstrates that analyzing connections or content in isolation is insufficient for modern threat detection. By fusing these dimensions, the framework can identify tiny, organized clusters that otherwise hide in the noise. Future work will focus on moving beyond detection to quantify the actual influence these groups exert on public opinion.

Summary updated

Abstract

Recent events have led to a burgeoning awareness on the misuse of social media sites to affect political events, sway public opinion, and confuse the voters. Such serious, hostile mass manipulation has motivated a large body of works on bots/troll detection and fake news detection, which mostly focus on classifying at the user level based on the content generated by the users. In this study, we jointly analyze the connections among the users, as well as the content generated by them to Spot Coordinated Groups (SCG), sets of users that are likely to be organized towards impacting the general discourse. Given their tiny size (relative to the whole data), detecting these groups is computationally hard. Our proposed method detects these tiny-clusters effectively and efficiently. We deploy our SCG method to summarize and explain the coordinated groups on Twitter around the 2019 Canadian Federal Elections, by analyzing over 60 thousand user accounts with 3.4 million followership connections, and 1.3 million unique hashtags in the content of their tweets. The users in the detected coordinated groups are over 4x more likely to get suspended, whereas the hashtags which characterize their creed are linked to misinformation campaigns.
BibTeX
@misc{wang2019sgp,
  author = {Junhao Wang and Sacha Levy and Ren Wang and Aayushi Kulshrestha and Reihaneh Rabbany},
  title = {{SGP: Spotting Groups Polluting the Online Political Discourse}},
  year = {2019},
  eprint = {1910.07130},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.07130},
}
Figure