layout: true class: center --- class: middle background-color: white .title[IRA Troll Accounts] .subtitle[Exploring their use and coordination] .author[Laila A. Wahedi, PhD] .date[November 29, 2018] .institution[MDI, McCourt School of Public Policy, Georgetown University]
.center[] .footnote[ Created with [{Remark.js}](http://remarkjs.com/) using [{Markdown}](https://daringfireball.net/projects/markdown/) + [{MathJax}](https://www.mathjax.org/) + [{Liminal}](http://www.jmlilly.net/liminal.zip) ] --- layout: true name: main class: center .footnote_left[ » Slides at https://wahedi.us under
Current Presentation
] .footnote_right[ Laila A. Wahedi • @lwahedi • law98@georgetown.edu ] --- template: main class: middle background-color: #78881f
Russian Trolls
What can we learn about Russian information operations by observing their activity online?
--- template: main background-color: #78881f
Russian Trolls
Data: ~3M tweets from ~3k known accounts from the Internet Research Agency
Released by FiveThirtyEight and
The Clemson Social Media Listening Center
Do have
Tweet text and time
Url redirects
Is quote or retweet
Don't have
Friends / Followers
Retweet user id
Interactions
--- template: main background-color: #78881f
What can we learn about how the IRA operated?
How were accounts managed?
Individual agents managing multiple accounts independently
Many accounts created for coordinated purposes
Expect:
Clusters of accounts with the same rhetoric/language
Larger clusters of accounts of accounts with similar rhetoric
--- template: main background-color: #01BAEF
Calculating Rhetorical Similarity: How?
1. Create Author Vectors
2. Calculate Distance
Weighted/ Thresholded
Relate authors by rhetoric
--- template: main background-color: #78881f
Russian Trolls
Large, loose clusters corresponding to category coding
No identifiable authors (small clusters with very similar author vectors
Suggests coordinated campaigns
--- template: main background-color: #78881f
Troll Coordination: Strategy
Mueller Indigtment: Goal was to infiltrate political movements to sew distrust in institutions
How did they do it?
Spy Strategy: Individually gain trust in existing networks?
Organizer Strategy: Jointly interact to create reputable network?
Expect to see...
Low levels of interaction among accounts.
High levels of interaction and mutual promotion.
--- template: main background-color: #78881f
Challenge: Measuring Promotion
Don't have
Friends / Followers
Retweet user id
Interactions
Three Approaches
Link Each Other
Link Same Content
Share Same Tweet (Retweet?)
Expect to see...
Dense, not strongly clustered
Dense, some clustering
Highly clustered by category
--- template: main background-color: #78881f
Link Each Other
As expected, links to tweets are not strongly clustered
Even negative content promotes tweets and credibility
--- template: main background-color: #78881f
Link Same Content
Surprisingly little clustering among mutual links
Surprisingly dense
--- template: main background-color: #78881f
Same Tweet (Retweets?)
Very strong clustering along type
Much smaller network
Core of mutually-promoting accounts, others excluded
--- template: main background-color: #78881f
To what extent were accounts coordinated?
Did trolls communicate a unified message?
How organized was their message?
If well coordinated, expect to see...
Trolls promoting similar message
Similar patterns of promotion over time
--- template: main background-color: #01BAEF
Topic Bipartite Network
1. Topic Model
2. Ties from user to topic
3. Optional: Project to unipartite
What makes authors similar?
How do topics relate?
--- template: main background-color: #78881f
Hashtags Over Time
Clusters by type link to clusters of hashtags
Coordinated use of hashtags over time
--- template: main background-color: #78881f
Unipartite Projection: Hashtags
Clear clustering of hashtag use by account type
Suggests coordinated messaging
--- template: main background-color: #F2DC5D
Questions?
Icons made by or adapted from Freepik from www.flaticon.com is licensed by CC 3.0