Detection of political manipulation in online communities through measures of effort and collaboration

Online social media allow users to interact with one another by sharing opinions, and these opinions have a critical impact on the way readers think and behave. Accordingly, an increasing number of manipulators deliberately spread messages to influence the public, often in an organized manner. In particular, political manipulation-manipulation of opponents to win political advantage-can result in serious consequences: antigovernment riots can break out, leading to candidates' defeat in an election. A few approaches have been proposed to detect such manipulation based on the level of social interaction (i.e., manipulators actively post opinions but infrequently befriend and reply to other users). However, several studies have shown that the interactions can be forged at a low cost and thus may not be effective measures of manipulation. To go one step further, we collect a dataset for real, large-scale political manipulation, which consists of opinions found on Internet forums. These opinions are divided into manipulators and nonmanipulators. Using this collection, we demonstrate that manipulators inevitably work hard, in teams, to quickly influence a large audience. With this in mind, it could be said that a high level of collaborative efforts strongly indicates manipulation. For example, a group of manipulators may jointly post numerous opinions with a consistent theme and selectively recommend the same, well-organized opinion to promote its rank. We show that the effort measures, when combined with a supervised learning algorithm, successfully identify greater than 95% of the manipulators. We believe that the proposed method will help system administrators to accurately detect manipulators in disguise, significantly decreasing the intensity of manipulation.