On the Influence of Twitter Trolls during the 2016 US Presidential Election

Nikos Salamanos, Michael J. Jensen, Xinlei He, Yang Chen, Michael SirivianosarXiv 2019 · OpenReview

Abstract: It is a widely accepted fact that state-sponsored Twitter accounts operated during the 2016 US presidential election spreading millions of tweets with misinformation and inflammatory political content. Whether these social media campaigns of the so-called "troll" accounts were able to manipulate public opinion is still in question. Here we aim to quantify the influence of troll accounts and the impact they had on Twitter by analyzing 152.5 million tweets from 9.9 million users, including 822 troll accounts. The data collected during the US election campaign, contain original troll tweets before they were deleted by Twitter. From these data, we constructed a very large interaction graph; a directed graph of 9.3 million nodes and 169.9 million edges. Recently, Twitter released datasets on the misinformation campaigns of 8,275 state-sponsored accounts linked to Russia, Iran and Venezuela as part of the investigation on the foreign interference in the 2016 US election. These data serve as ground-truth identifier of troll users in our dataset. Using graph analysis techniques we qualify the diffusion cascades of web and media context that have been shared by the troll accounts. We present strong evidence that authentic users were the source of the viral cascades. Although the trolls were participating in the viral cascades, they did not have a leading role in them and only four troll accounts were truly influential.

Introduction

Since the 2016 US Presidential Election, there has been a heated debate over whether state-sponsored accounts (primarily from Russia, Iran, and Venezuela) actually moved the needle of public opinion. Previous studies faced significant hurdles: Twitter's API only provides a 1% sample, deactivated accounts disappear from history, and social graphs are often private. This study set out to quantify the actual influence of these trolls—were they the leaders of the digital mob, or were they just shouting into a void?

Methodology

The researchers employed a multi-pronged graph-theoretical and machine-learning approach to reconstruct the interaction networks and information cascades of the 2016 election period:

  • Interaction Graph Construction: A massive directed multigraph was built where nodes represent users and edges represent "actions" (replies and mentions), serving as a proxy for the social graph.
  • Causality-Based Diffusion Cascades: To track information spread, a "time-inferred" model was used. If User A shared a URL and User B (who interacts with A) shared it later, the model assumes a causal link.
  • Centrality Metrics: The researchers used k-core decomposition to identify the "inner core" of the network—the most influential and well-connected spreaders.
  • Bi-LSTM Classification: To ensure the influential "authentic" users weren't just trolls in disguise, a Bidirectional Long Short-Term Memory (Bi-LSTM) model was trained using Doc2Vec embeddings to compare the behavior of trolls versus real users.

Datasets

The study relies on two massive pillars of data:

  1. The Live Stream: 152.5 million tweets from 9.9 million users collected in real-time between Sept 21 and Nov 7, 2016. Crucially, this includes "pre-deletion" tweets from trolls.
  2. The Ground Truth: Twitter’s official 2018 release of 8,275 state-sponsored accounts. By cross-referencing these, the authors identified 822 active trolls within their own live-streamed dataset.

Key Insights and Ideas

The Power of the "Ego-Net"

The most surprising finding is that the trolls' "ego-net"—the authentic users who interacted with them—were the real influencers. These users were more popular, more social, and more central to the network than the trolls themselves.

Figure 6: CCDF of out-degree for spreaders in the simple graph, showing authentic users in the ego-net were significantly more active than trolls.

Trolls as Bandwagoners

Rather than injecting brand-new viral content, trolls appeared to "resonate" with existing communities. They targeted opinion leaders and joined existing cascades. In fact, 85% of influential authentic users appeared in cascades before the trolls did.

The "Four Horsemen"

Despite the thousands of accounts involved, influence was extremely top-heavy. Only four troll accounts were found to be truly influential in terms of driving real users to spread their content. When the researchers mathematically "removed" the trolls from the cascades, the overall structural virality of the information didn't change significantly, suggesting the misinformation would have likely spread with or without them.

Key Results

The analysis of the 9.3 million node interaction graph revealed a significant popularity gap between trolls and authentic users. While 5,223 authentic users had an "in-degree" (popularity) greater than 1,000, only 12 trolls reached this threshold. Furthermore, authentic users in the trolls' ego-nets initiated over 54,000 cascades, while trolls themselves only initiated 647.

Figure 3: CCDF of in-degree for spreaders in the multigraph, illustrating that authentic users (ego-net) consistently possessed higher popularity than trolls.

Limitations and Conclusion

The authors acknowledge that using the 1% Twitter Streaming API is a limitation, as it may miss smaller cascades. Additionally, because the interaction graph was built on mentions and replies, it remains an approximation of the true social network. However, the study provides strong evidence that while state-sponsored trolls were prolific, they were rarely the "patient zero" of viral misinformation. Instead, a small group of highly active, authentic users acted as the primary engine for virality, with trolls often just bandwagoning on existing trends to gain resonance.

Summary updated

Abstract

It is a widely accepted fact that state-sponsored Twitter accounts operated during the 2016 US presidential election spreading millions of tweets with misinformation and inflammatory political content. Whether these social media campaigns of the so-called "troll" accounts were able to manipulate public opinion is still in question. Here we aim to quantify the influence of troll accounts and the impact they had on Twitter by analyzing 152.5 million tweets from 9.9 million users, including 822 troll accounts. The data collected during the US election campaign, contain original troll tweets before they were deleted by Twitter. From these data, we constructed a very large interaction graph; a directed graph of 9.3 million nodes and 169.9 million edges. Recently, Twitter released datasets on the misinformation campaigns of 8,275 state-sponsored accounts linked to Russia, Iran and Venezuela as part of the investigation on the foreign interference in the 2016 US election. These data serve as ground-truth identifier of troll users in our dataset. Using graph analysis techniques we qualify the diffusion cascades of web and media context that have been shared by the troll accounts. We present strong evidence that authentic users were the source of the viral cascades. Although the trolls were participating in the viral cascades, they did not have a leading role in them and only four troll accounts were truly influential.
BibTeX
@misc{salamanos2019influence,
  author = {Nikos Salamanos and Michael J. Jensen and Xinlei He and Yang Chen and Michael Sirivianos},
  title = {{On the Influence of Twitter Trolls during the 2016 US Presidential Election}},
  year = {2019},
  eprint = {1910.00531},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.00531},
}
Figure