Contributions
- Dataset and methodology: Collected election-related tweets from Sept 16–Oct 21, 2016, using a curated list of 23 hashtags/keywords balanced across candidates. Demonstrates scalable, network-based user ideology classification via label propagation.
- Quantification of asymmetric amplification: Documents a striking 31:1 disparity in conservative vs. liberal retweets of Russian trolls, and a 36:1 disparity in volume of amplifying tweets, despite roughly equal troll numbers in each ideological camp.
- Bot activity profiling: Applies Botometer at scale to classify spreaders; finds bots present across the ideological spectrum (~4.9% liberal, ~6.2% conservative) but concentrated in conservative amplification chains.
- Geospatial and textual evidence: Maps Russian troll engagement across U.S. states (Tennessee and Texas highest) and documents conservative keywords (e.g., #trumpforpresident, #MAGA, #draintheswamp) dominating troll content.
Method
The paper constructs a retweet network of 4.7M nodes and uses label propagation to infer user ideology. Seeds are 29k users who tweeted links to partisan media outlets (Huffington Post, Breitbart, etc.); their ideological labels are propagated through the retweet graph via a standard label-propagation algorithm with in-degree weighting. Cross-validation on held-out seed users and hyper-partisan Twitter accounts (those with partisan outlet links in their profile URL) yields precision and recall both ~0.91–0.93.
For bot detection, the authors use the Botometer framework (Davis et al. 2016), which extracts >1000 features spanning account metadata, profile customization, geolocation, activity statistics, and temporal patterns. A score >0.5 indicates a bot. Bots are most reliable signals for: default profile setting, absence of location metadata, high tweet frequency, high retweet-to-original ratio, and recently-created accounts.
Russian trolls are identified from a U.S. Congress-released list of 2,752 accounts tied to the Internet Research Agency (IRA). Of these, 221 appear in the dataset; 85 wrote original content (861 tweets total). Users who retweeted trolls are classified as "spreaders."
Results
Russian troll content: Although predicted ideology is roughly split (107 liberal, 108 conservative trolls), production is heavily skewed: only 15 liberal trolls wrote originals (44 tweets), vs. 64 conservative trolls (844 tweets). Top conservative keywords include #trumpforpresident (486 occurrences), #MAGA (113), #neverhillary (106). Liberal trolls rarely produced original content.
Spreader (amplifier) behavior: 40,224 distinct spreaders retweeted trolls. Among those writing original tweets (28,274 spreaders), conservatives vastly outpaced liberals: 27,382 conservative spreaders vs. 892 liberal spreaders (31:1 ratio); 1.5M+ conservative tweets vs. 42k liberal tweets (36:1 ratio).
Bot presence: Of 34,160 spreaders with bot scores, 75 of 1,506 liberals (4.9%) and 2,018 of 32,513 conservatives (6.2%) scored as bots. Absolute numbers are dominated by conservatives due to the 22:1 conservative-to-liberal spreader ratio. Bot activity (tweets per bot) is similar across ideologies (~8%).
Geography: Geolocation analysis (from self-reported profile locations and geotagged tweets) shows liberal spreader tweets concentrated in Democratic states (New York prominent), while conservative spreader tweets distributed across more states and concentrated in Republican strongholds (Texas, Florida).
Notes
Strengths: The scale of the dataset (43.7M tweets), the high-precision ideology classification (>0.91 precision and recall on held-out validation), and the clear quantification of asymmetric amplification make this a robust empirical foundation. The paper is methodologically sound: label propagation on retweet networks is well-established for ideology inference, and cross-validation on both seed users and hyper-partisan accounts strengthens confidence in the labeling. The combination of text, network, bot, and geospatial analysis provides multiple angles on the same phenomenon.
Limitations and follow-ups: The paper uses a snapshot (Sept 16–Oct 21, 2016) and cannot capture the full scope of IRA operations; later work has shown the IRA was active well before and after the election. The label propagation seeds come only from the top 5 liberal and 5 conservative outlets, which may miss ideologically moderate or fringe users. Bot detection thresholds (Botometer >0.5) are coarse; the paper notes that scores >0.5 are conservative but some bots may fall below it. The paper does not directly measure the causal impact of Russian troll content on election outcomes—only that conservatives amplified it more, a necessary but not sufficient condition for influence. The self-reported geolocation analysis includes substantial noise (fictitious or imprecise locations), mitigated somewhat by the large scale but worth noting.
Methodological significance: This work has been foundational in popularizing retweet-network-based ideology classification at scale and demonstrating the application of bot-detection tools (Botometer) to large, heterogeneous Twitter populations in a real-world political context. The asymmetry finding—that a single troll farm could engineer a 31-fold disparity in amplification—is a key empirical result in understanding how coordinated disinformation campaigns exploit existing ideological polarization.