PaperlayerView exact revision on arXiv

Source-linked AI summary

Characterizing the 2016 Russian IRA Influence Campaign

Adam Badawy, Aseel Addawood, Kristina Lerman, Emilio Ferrara

arXiv:1812.01997v1cs.SI

TL;DR

The paper asks how Russian troll accounts’ content spread through Twitter’s 2016 election-related discussion. It analyzes users who engaged with the accounts, finding greater sharing by conservatives and bots among both liberal and conservative users, while noting that the analysis cannot capture all manipulation efforts.

  • Problem

    The study examines the effects of Russian troll activity on election-related political discussion, including who shares the trolls’ content and how it spreads.

    Source: §1 · p. 3, §1.1 · p. 3

  • Method

    The authors analyze Russian troll content and its spread, using label propagation for ideology and examining retweet-network positions and bots.

    Source: §1 · p. 3, §1.1 · p. 3

  • Results

    Conservatives shared Russian troll propaganda more widely than liberals, and bots appeared among both liberal and conservative users.

    Source: §1.2 · p. 3, §5 · p. 9

  • Takeaways & Limitations

    The analysis describes ideological, network, and bot patterns among users who engaged with Russian troll content in 2016.

    Source: §1.1 · p. 3, §5 · p. 9

    The analysis cannot account for all malicious efforts aimed at manipulation during the 2016 election.

    Source: §5 · p. 9

Abstract

from arXiv · show

Until recently, social media were seen to promote democratic discourse on social and political issues. However, this powerful communication ecosystem has come under scrutiny for allowing hostile actors to exploit online discussions in an attempt to manipulate public opinion. A case in point is the ongoing U.S. Congress investigation of Russian interference in the 2016 U.S. election campaign, with Russia accused of, among other things, using trolls (malicious accounts created for the purpose of manipulation) and bots (automated accounts) to spread propaganda and politically biased information. In this study, we explore the effects of this manipulation campaign, taking a closer look at users who re-shared the posts produced on Twitter by the Russian troll accounts publicly disclosed by U.S. Congress investigation. We collected a dataset of 13 million election-related posts shared on Twitter in the year of 2016 by over a million distinct users. This dataset includes accounts associated with the identified Russian trolls as well as users sharing posts in the same time period on a variety of topics around the 2016 elections. We use label propagation to infer the users' ideology based on the news sources they share. We are able to classify a large number of users as liberal or conservative with precision and recall above 84%. Conservative users who retweet Russian trolls produced significantly more tweets than liberal ones, about 8 times as many in terms of tweets. Additionally, trolls' position in the retweet network is stable over time, unlike users who retweet them who form the core of the election-related retweet network by the end of 2016. Using state-of-the-art bot detection techniques, we estimate that about 5% and 11% of liberal and conservative users are bots, respectively.

1 Introduction

The introduction frames social media manipulation as a threat to political discussion and studies the spread of content from identified Russian troll accounts. It outlines analyses of ideology, retweet-network positions, bots, and geography, alongside the paper’s main contributions.

1 Introduction

  • The study examines how content from accounts identified as Russian trolls spread on Twitter and affects election-related political discussion.The authors focus on user intent to deceive rather than deciding which information qualifies as fake news.Source: §1 · pp. 2–3

1.1 Research Questions

  • The research questions cover ideological engagement, trolls’ and spreaders’ network positions over time, bots’ role, and geographic variation in propagation.The geographic analysis tests whether users in particular states participate in consuming and spreading troll content more than others.Source: §1.1 · p. 3
  • The study extends prior work from two months to a year of data and adds network analysis of malicious users over time.The dataset contains more than 13 million tweets from over a million users.Source: §1.1 · p. 3

1.2 Summary of Contributions

  • Political-ideology classification achieves precision and recall above 84% for most users in the dataset.The paper uses network-based machine learning methods for the classification.Source: §1.2 · p. 3
  • Text analysis finds conservative trolls focused on refugees, terrorism, and Islam, while liberal trolls focused on school shootings and police.Trump and Clinton also appear among the top words used by both groups.Source: §1.2 · p. 3
  • The geospatial analysis finds that some states were unusually engaged in producing and diffusing Russian troll content.The introduction also reports bots among users engaging with trolls in both liberal and conservative domains.Source: §1.2 · p. 3

2 Data Collection

The study assembled Twitter data from identified Russian troll accounts and election-related users, then established political-leaning media outlets as a basis for labeling users. It also described the dataset’s activity over 2016.

2.1.1 Trolls

  • 1,148 accounts from the identified troll list appear in the dataset, and just over a thousand produced more than half a million original tweets.The study used a list of 2,752 accounts identified as Russian trolls.Source: §2.1.1 · pp. 3–4
  • Non-troll data included election-keyword tweets and other tweets from the same users that omitted those keywords.The keyword list covered major presidential candidates, debate terms, and third-party candidates.Source: §2.1.1 · p. 4
  • Tweets and users were tracked across 2016, with a spike in activity around the election.More than three-quarters of the dataset consisted of retweets and tweets with URLs.Source: §2.1.1 · p. 4, Figure 1 · p. 4

2.2 Classification of Media Outlets

  • American Thinker dominated the conservative sample’s shared URLs, while liberal-sample URLs were spread more evenly across outlets.The outlet domains were listed after removing left-center and right-center sources.Source: §2.2 · p. 5, Table 3 · p. 5
  • The researchers combined liberal with liberal-center outlets and conservative with conservative-center outlets, then labeled users by which side’s linked sources they shared more often.The outlet lists contained 641 liberal and 398 conservative sources; the final labeled set contained 10,074 users.Source: §2.2 · pp. 4–5

3 Methods and Data Analysis

The methods use a retweet network to propagate ideology labels, detect bots among active users, and estimate tweet locations. Seed-label validation yielded precision and recall around 0.84, while the location data include millions of tweets.

3.1 Retweet Network

  • The directed retweet network links users when one retweets the other and has a giant component of 1,357,493 nodes.The network is described as sparse.Source: §3.1 · p. 5

3.2 Label Propagation

  • Label propagation iteratively assigns each node its most frequent neighbors’ label, using fixed seed labels and retweet counts as weights.The study seeded the network with more than 10,000 users labeled from the media outlets they mainly retweeted.Source: §3.2 · p. 5
  • Precision and recall were around 0.84 in stratified five-fold validation of more than 10,000 ideology-labeled seeds.The validation combined liberal with liberal-center labels and conservative with conservative-center labels, including users near the political center.Source: §3.2 · p. 5, Table 5 · p. 5

3.3 Bot Detection

  • Botometer analyzes over 1,000 features across six subclasses, and accounts scoring above 0.5 are labeled bots.The Python API uses 300 recent tweets and publicly available account metadata to produce bot scores.Source: §3.3 · p. 6

3.4 Geolocation

  • Crimson Hexagon combines available geotags with contextual estimates of users’ locations; the dataset includes over 9 million country-located tweets.More than 7.5 million were from the United States, and over 4.7 million U.S. tweets had state information.Source: §3.4 · pp. 6–7

4 Results

The results compare how political ideology, network position, bot activity, and location relate to sharing Russian trolls’ content. Spreaders become increasingly central over time, and bot scores and state-level retweet proportions differ across political groups.

4.1 RQ1: Political Ideology

  • Conservative spreaders were fewer but wrote substantially more tweets than liberal spreaders, while liberal trolls were fewer yet produced more tweets than conservative trolls.The label-propagation method labeled most users, though it could not label some users outside the giant component.Source: §4.1 · p. 7, Table 7 · p. 7, Table 6 · p. 7

4.2 RQ2: Temporal Analysis

  • Spreaders progressively increased their centrality and dominated the retweet network’s core toward year-end, while trolls remained concentrated in lower cores and fairly stable over time.Monthly network snapshots were additive, retaining nodes and edges from preceding months.Source: §4.2 · p. 8, Figure 3 · p. 9

4.3 RQ3: Social Bots

  • About 5% of liberal spreaders and 11% of conservative spreaders with bot scores exceeded 0.5; conservative spreaders also had higher mean bot scores, 0.29 versus 0.18.Conservative spreaders had higher means across all Botometer subclass scores; all differences were statistically significant (p <0.001).Source: §4.3 · p. 8, Figure 4 · p. 9

4.4 RQ4: Geospatial Analysis

  • Conservative users’ state-level retweet proportions averaged 0.34 (SD 0.12), versus 0.26 (SD 0.13) for liberals; Wyoming led for conservatives, while Montana and Kansas were significant liberal outliers.The proportions normalize troll retweets by each state’s total tweets for the corresponding ideology.Source: §4.4 · p. 8

5 Conclusions

The paper examines Russian trolls’ Twitter content as a proxy for propaganda and finds it was shared more widely by conservatives. It also notes that the analysis cannot account for all malicious efforts to manipulate the election.

5 Conclusions

  • Russian-troll content was shared more widely by conservatives; conservative spreaders wrote about 3.2 times as many tweets as liberal spreaders, despite being fewer.The estimated bot shares were about 5% among liberal users and 11% among conservative users.Source: §5 · p. 9
  • The study uses Russian trolls’ Twitter content as a proxy for propaganda and analyzes tweets from trolls, spreaders, and other users during 2016.Source: §5 · p. 9
  • The analysis cannot account for all malicious efforts aimed at manipulation during the 2016 presidential election.The authors call for more sophisticated techniques to detect and counter such efforts.Source: §5 · pp. 9–10

Related papers

Loading 1812.01997v1…