Fig 1.
Diagram of the different steps taken in our analysis of bots and disinformation in the Twitter discussion surrounding Donald Trump’s first impeachment.
Table 1.
Data collection results, by account type.
Fig 2.
Bot prevalence and activity rates.
(left) Fraction of bots and (right) average tweet rate of each account type category.
Fig 3.
Bot language and content analysis.
(left) Average media quality score and (right) average toxicity score of tweets posted by different account types.
Table 2.
Statuses most retweeted by anti-Trump bots.
Table 3.
Statuses most retweeted by pro-Trump bots.
Fig 4.
Number of retweets by (top) anti-Trump bots and (bottom) pro-Trump bots for top retweeted Twitter accounts.
Fig 5.
Bot follower network reach and composition.
(left) Venn diagram of users who follow anti-Trump and pro-Trump bots. (right) Fraction of co-partisan followers for different bot types.
Fig 6.
Bot follower network structure.
(left) Follower network of bot accounts colored by partisanship, where anti-Trump bots are blue, and pro-Trump bots are red. (right) Follower network of Qanon accounts, where bots are colored red and humans are colored green.
Fig 7.
(top) Daily generalized harmonic influence centrality (GHIC) score for all bots versus date. (bottom) Boxplot of the daily GHIC per bot for different bot types.
Table 4.
Tweet collection terms.
Table 5.
Keywords used to identify anti-Trump users.
Table 6.
Keywords used to identify pro-Trump users.
Fig 8.
Confusion matrix of trained partisanship classifier applied to test data.
Fig 9.
Histogram of the partisanship scores of all users.
Table 7.
Example partisanship classification scores.
Table 8.
Q-anon user profile terms.
Fig 10.
Bot probability classifications on an example day.
Histogram of the bot probabilities calculated by the factor graph algorithm [40] based on the impeachment retweet network on February 1, 2020 (an example day).
Fig 11.
ROC curves for (left) “overall” Botometer scores and (right) “astroturf” Botometer scores using our bot classification results as ground truth.
Fig 12.
Violin plot of the mean user toxicity score for different user types.
Fig 13.
Daily generalized harmonic influence centrality (GHIC) score for all bots versus date.
The error bars correspond to the range of GHIC values for bot probability thresholds of 0.72, 0.80, and 0.88.