Who Falls for Online Political Manipulation?

1 INTRODUCTION

• The study examines Russian trolls’ role in the 2016 US election amid concerns about misinformation and bots in political discourse.Prior reports on misinformation and bots increased interest in automatically detecting and predicting malicious actor activity.Source: §1 · p. 1
• The study defines trolls as users intending to deceive or create conflict and asks whether users who spread their messages can be predicted and distinguished from others.The authors call these users “spreaders” and frame prediction as a possible basis for a counter-campaign.Source: §1 · pp. 1–2
• The researchers collected election-related Twitter posts from September 16 to November 9, 2016, identified 221 congressional-listed troll accounts, and used user features in machine-learning models.The dataset contained over 43 million posts from about 5.7 million users.Source: §1 · p. 2
• 96% average AUC was achieved with Gradient Boosting on the roughly balanced subset under 10-fold validation; the full dataset also exceeded 90%.Political ideology was the most prominent feature, followed by followers, tweet counts, and bot scores among the top predictors.Source: §1 · p. 2

2 RELATED LITERATURE

• Reports describe online manipulation in US elections, South America [Forelle et al. 2015], the U.K. [Howard and Kollanyi 2016], and Italy [Cresci et al. 2017].Researchers had warned of online political manipulation for over a decade.Source: §2 · p. 2
• Bots have played substantial roles in political manipulation; one study attributed 3.8 million election-month tweets to 400,000 bots, one-fifth of its collected conversation [Bessi and Ferrara 2016].Russian interference was also reported in Germany [Applebaum and Colliver. 2017], British elections, and the Catalonian referendum [Stella et al. 2018].Source: §2 · p. 3
• Similar efforts are reported beyond Russia, including Chinese online distraction campaigns and Korean bots supporting an incumbent candidate.Russian government-affiliated trolls also participated in domestic Russian online discussions and Wikipedia editing [Ananyev and Sobolev 2017; Labzina 2017].Source: §2 · p. 3
• Although fake-news stories were widely shared in 2016, an average American saw only a few; studies also find that accurate political information can change views.Research links misinformation and limited political knowledge with differences in individual and collective opinion.Source: §2 · p. 3
• Political misperceptions can distort collective opinion and policy debates; trolls used misinformation about immigration, minority issues, and government to advance campaign agendas.The literature defines directionally motivated reasoning as selectively accepting or rejecting information according to prior beliefs and attitudes.Source: §2 · pp. 3–4

3.1 Twitter Dataset

• Continuous, uninterrupted Twitter Search API queries collected 43.7 million unique tweets from nearly 5.7 million users between September 15 and November 9, 2016.The researchers selected 23 election-related hashtags and keywords across major candidates and debates, and chose Search API collection to avoid known sampling issues with the Stream API [Morstatter et al. 2013].Source: §3.1 · p. 4

3.2 Russian Trolls

• Of 2,752 congressional-identified troll accounts, 221 appeared in the dataset; 85 produced 861 original tweets, and the trolls were retweeted 83,719 times.Three troll accounts received over 89% of those retweets, and 40,224 distinct users retweeted trolls.Source: §3.2 · pp. 4–5, Table 2 · p. 5

3.3 Spreaders

• Among roughly 40,000 spreaders, 28,274 produced original tweets; they generated over 1.5 million original tweets and over 12 million other tweets and retweets.Spreaders are users who rebroadcast content produced by Russian trolls.Source: §3.3 · p. 5, Table 3 · p. 5

4 DATA ANALYSIS & METHODS

• Features span Metadata, LIWC, Engagement, Activity, and Other; most are not highly correlated, while engagement measures show a “rich get richer” pattern.Word Count and Tweet Count, Positive Emotion and Affection, and Anxiety and Anger are highly correlated pairs.Source: §4 · p. 6, Table 4 · p. 6

4.1.1 Classification of Media Outlets.

• Users were labeled liberal or conservative by which side’s media links they shared more often; users with equal counts were removed, leaving 29,832 labeled users.The outlet lists drew on five frequently appearing outlets per partisan category.Source: §4.1.1 · pp. 6–8

4.1.2 Label Propagation.

• A retweet network propagates liberal or conservative labels from 29K seed users selected through partisan media outlets, whose seed labels remain fixed.Label propagation iteratively assigns each node the most frequent label among its neighbors.Source: §4.1.2 · p. 8
• Seed-user validation averaged over 91% precision and recall; hyper-partisan-user test precision and recall exceeded 93%.Table 7 reports that both validation methods labeled over 90% of test-set users correctly.Source: Table 7 · p. 8, §4.1.2 · p. 8

4.2 Bot Detection

• Botometer scores were obtained for over 34K spreaders and a roughly equivalent random sample of circa 37K non-spreaders; scores above 0.5 marked bots.The spreaders’ mean score was 0.3, higher than the non-spreaders’ mean.Source: §4.2 · p. 9
• Spreaders differed on almost all Botometer subclass scores except temporal features; their means were higher across subclasses, with metadata, friends, and network distributions also substantively different.The reported differences were statistically significant at p < 0.001.Source: §4.2 · p. 9

4.3 Engagement

• Engagement measures cover retweets, mentions, replies, and quotes through quantity, longevity, and stability, including an h-index for sustained referencing over time.The h-index captures how highly and continuously a user is referenced by others in the network.Source: §4.3 · pp. 9–10

• The balanced dataset contained about 72,000 users, and the study compared four off-the-shelf classifiers using stratified 10-fold cross-validation.Source: §5 · p. 10
• AUC rose from 89% in Model 4 to 96% in Model 5, with bot scores and political ideology added; Figure 4 compares model AUCs.Source: §5 · p. 11, Figure 4 · p. 12
• Political ideology ranked first in predictive importance, followed by followers count, statuses count, and bot score; Figure 5 reports the feature ranking.Source: §5 · p. 12, Figure 5 · p. 13
• Partial dependence showed positive relationships with political ideology, statuses count, bot scores, and friends count, and negative relationships with followers count, account age, character count, and word count.Source: §5 · p. 13, Figure 6 · p. 14
• AUC ranged from 83% for the baseline model to 98% for the full model on the 5.7-million-user dataset.Source: §5 · p. 13

6 DISCUSSION AND LIMITATIONS

• Political ideology was highly predictive of who spread trolls’ messages, while basic account metadata and bot scores also distinguished spreaders from non-spreaders.Botometer friend, network, and user-metadata subclasses all showed higher bot scores for spreaders.Source: §6 · p. 14
• Spreaders tended to tweet more, post less original content, have higher bot scores, be more conservative, have fewer followers, and hold newer accounts.Political ideology had the most influence among the plotted features; follower count and account age were more influential than other features in Figure 6b.Source: §6 · p. 14
• The dataset omitted some listed trolls, and the authors could not gauge how that missingness affected their results.Some absent accounts may simply have been inactive during data collection.Source: §6 · p. 14
• Generalization is constrained by uncertainty about how the troll list was compiled, conservative-targeting skew, possible tweet-length limits for LIWC categories, and election-specific circumstances.The authors say the conservative-targeting result offers little insight into liberals’ responses; the same phenomenon may or may not unfold similarly elsewhere.Source: §6 · p. 15
• Despite these limitations, the authors used computational tools to analyze and predict troll activity in the 2016 election and to begin examining whether manipulation can be stopped or prevented.Their keyword-collected dataset was intended to encompass the political event under study.Source: §6 · p. 15

7 CONCLUSION

• Identifying likely victims of political manipulation is presented as a first step toward containing the spread of malicious content.The authors argue that social media platforms should attempt to contain this spread.Source: §7 · p. 15
• Individual misperceptions can aggregate into distorted collective public opinion, with potentially severe policy implications.Source: §7 · p. 15
• Declining trust in mainstream news and increased exposure to content from ill-intended sources are described as dangers to democratic life.Source: §7 · p. 15