[Paper Review] Who Falls for Online Political Manipulation?

Adam Badawy, Kristina Lerman+1 more|arXiv (Cornell University)|Aug 9, 2018

Misinformation and Its Impacts4 citations

TL;DR

This study develops a machine learning model to predict users susceptible to spreading Russian troll content during the 2016 U.S. election, using Twitter data. It achieves 96% AUC in identifying spreaders by leveraging political ideology, bot likelihood scores, and user activity metrics, demonstrating that these features are highly predictive of vulnerability to online political manipulation.

ABSTRACT

Social media, once hailed as a vehicle for democratization and the promotion of positive social change across the globe, are under attack for becoming a tool of political manipulation and spread of disinformation. A case in point is the alleged use of trolls by Russia to spread malicious content in Western elections. This paper examines the Russian interference campaign in the 2016 US presidential election on Twitter. Our aim is twofold: first, we test whether predicting users who spread trolls' content is feasible in order to gain insight on how to contain their influence in the future; second, we identify features that are most predictive of users who either intentionally or unintentionally play a vital role in spreading this malicious content. We collected a dataset with over 43 million elections-related posts shared on Twitter between September 16 and November 9, 2016, by about 5.7 million users. This dataset includes accounts associated with the Russian trolls identified by the US Congress. Proposed models are able to very accurately identify users who spread the trolls' content (average AUC score of 96%, using 10-fold validation). We show that political ideology, bot likelihood scores, and some activity-related account meta data are the most predictive features of whether a user spreads trolls' content or not.

Motivation & Objective

• To investigate whether users who spread Russian troll content can be predicted before they do so.
• To identify the most predictive features distinguishing users who intentionally or unintentionally spread malicious political content.
• To inform the development of early-warning systems to mitigate the spread of disinformation on social media.
• To understand the role of political ideology, bot behavior, and user activity in susceptibility to manipulation.

Proposed method

• Collected 43 million election-related tweets from 5.7 million users between September 16 and November 9, 2016, using keyword- and hashtag-based Twitter API queries.
• Cross-referenced the dataset with the U.S. Congress’s list of 221 Russian troll accounts to identify content they posted.
• Trained multiple machine learning classifiers, including Gradient Boosting, on user-level features such as political ideology, follower count, tweet volume, and bot likelihood scores.
• Used 10-fold cross-validation to evaluate model performance, with Area Under the Curve (AUC) as the primary metric.
• Conducted partial dependence analysis to interpret feature importance and model behavior, particularly for political ideology and bot scores.
• Validated results on both balanced subsets and the full dataset, including cases with missing values, to ensure robustness.

Experimental results

Research questions

• RQ1 Can we predict which users will become susceptible to spreading content from Russian trolls during the 2016 U.S. election?
• RQ2 What user-level features most strongly predict whether a user will spread content from Russian trolls?
• RQ3 How do political ideology, bot behavior, and user activity levels influence susceptibility to manipulation?
• RQ4 To what extent can machine learning models distinguish between users who spread troll content and those who do not?

Key findings

• The model achieved an average AUC score of 96% in distinguishing between users who spread Russian troll content and those who did not, using the full set of features and Gradient Boosting.
• Political ideology was the most predictive feature, with conservative-leaning users showing higher susceptibility to spreading troll content.
• Bot likelihood scores and user activity metrics—such as number of tweets and follower count—were among the top predictive features.
• Users who spread troll content tended to be highly active (posting many tweets), have high bot scores, and be more conservative, though they often had fewer original posts and fewer followers.
• The model maintained high performance (over 90% AUC) even on the full dataset and when excluding rows with missing values, indicating robustness.
• Partial dependence plots confirmed that political ideology had the strongest influence on the model’s prediction, followed by follower count and account age.