How rumors spread in online social media

Investigating large–scale Twitter data reveals that rumors spread in a drastically different temporal, structural, and linguistic pattern from non-rumors.

Article  |  Fall 2015

Detecting rumors has been an important problem in research as well as in everyday life because certain rumors that spread widely can cause critical damage to a person or even to a wider community. Until recently, the key social and human behavioral patterns related to rumor propagations remained largely as theories and hypotheses due to lack of data to observe them. The advent of social media has provided a new opportunity to examine how rumors propagate in life.

Ph.D. student Sejeong Kwon and Dr. Meeyoung Cha from the Graduate School of Culture Technology in KAIST investigated rumor propagation patterns based on large-scale data. The researchers examined millions of tweets related to over 100 viral propagation events on Twitter and identified the unique temporal, structural, and linguistic characteristics of rumors.

Rumors diffused in a different fashion from non-rumors. The most striking signal was found in temporal trends; rumors exhibited more fluctuations and shorter cycles over time, which meant that rumors were promoted more frequently than other types of information and were less likely to incur any chain-like conversations. The researchers proposed a new temporal model called PES (Periodic External Shocks) to demonstrate the cyclic behaviors of rumors.

Another distinguishing signal was seen in the network structure of those users participating in rumor propagation. Rumors spread through low degree users in Twitter, who have fewer followers than others. As a result, rumor propagation networks had sparse connections and contained many more singletons than non-rumor networks. Figure 2 represents an example in which a rumor that claimed someone saw the legendary creature Bigfoot shows a vastly different network shape from a non-rumor that mentioned that Twitter will buy the IT company Summize. Furthermore, rumors exhibited certain linguistic features such as negation and speculation, which could be detected automatically with sentiment analysis tools. When combined, a machine-learning algorithm could identify rumors from non-rumors with nearly 90% accuracy.

The study provides insights into why and how people participate in rumor spreading based on big data analysis, and its novel rumor detection features could play an important role in society. The study is also an excellent example of an interdisciplinary research, as it goes beyond data analysis and explains findings in terms of temporal, structural, and linguistic characteristics with existing social and psychological theories about rumor spreading.

This project was conducted in collaboration with Dr. Kyomin Jung of Seoul National University and Dr. Wei Chen and Dr. Yajun Wang of Microsoft Research Asia. This study was presented at the IEEE International Conference on Data Mining (ICDM) in 2013 and was featured in prestigious media outlets like New Scientist and the Washington Post.

[ICDM 2013]
http://icdm2013.rutgers.edu/schedule

Additional links for more information:

New Scientist
http://tinyurl.com/qcyffl7

Washington Post
http://tinyurl.com/papjjgf