Detecting fake news in social networks through crowd signals
Link: https://dl.acm.org/citation.cfm?id=3188722
The propagation of social media ushered in an era where information and news could be disseminated almost instantly. While there are undoubtedly positive impacts from this, the 2016 U.S. Presidential Election is an illustration of the weaponization of fake news. In this academic journal, the researchers sought to utilize crowd signals (Facebook’s implementation of a flagging tool for users) in determining the bad nodes within a network. Similar to the systems in place to detect email and SMS spam, the researchers go a step further in the creation of their own system: prior research in determining a user’s reputation (and weighting the results) assumes that the reputation is known to the system, and their algorithm (“Detective”) was built to learn about users’ flagging behavior while simultaneously detecting fake news.
Their setup consisted of 4,039 nodes/users and 88,234 edges (collected from survey data through the Facebook app), of which 20% of users generate fake news with a probability of .6, 40% generate fake news with probability of .2, and the remaining 40% generate fake news with probability of .01. Their experimental results indicate that Detective’s performance increased to the level of a fictional, model algorithm as user parameters were processed- in addition, Detective was still effective when used with a user base that was primarily adversarial/composed of spammers.
Akin to the applications of networks as discussed in class, particularly the behavior of nodes within a system, the interactions between nodes are vastly important to future models. The study found that very few users can reliably flag fake news, which leads me to question how networks could be applied to teach other users to more reliably flag fake news. Behavior in this context is learned, and the insights that we gain from studying networks to accomplish “x” could double as a way to teach users about accomplishing “x” on their own. In other words, addressing the root cause of the problem (users unable to accurately gauge fake news) in synthetically bringing them together with the “right” nodes (users able to consistently and accurately gauge fake news). Just a thought.