Studying Fake News via Network Analysis: Detection and Mitigation
Recent political drama has shown the effects of fake news and it’s impact on social media networks. Shu, et al studies the detection of fake news via different network properties as well as mitigation techniques via network analysis. The cornerstone of the study is in a news dissemination framework that identifies valuable features within networks that are indicative of fake news.
The first component of the framework is the content component, referred to as the content dimension in the paper. The content dimension focuses on the tangible content within the fake media. For example, the content in a fake news Facebook post can be used as a very high level feature to identify fake content. Shu, et al showcase that deep learning neural networks can be used to perform a classification task on the content dimension. Specifically, the content in a piece of fake news media is very high dimensional, there can be multiple features within the content that are indicative of fake vs. real news. Thus the need for large, high dimensional mappings of content to classifications (real vs. fake).
The second component of the framework is the social component. Specifically, who is posting the fake news content. During a network analysis, we can identify individuals that fall into three different categories. The first are persuaders: individuals that populate and circulate fake news media. The second are gullible users: individuals that are credulous and can be easily persuaded to believe in the fake content. The third are clarifiers: individuals that propose skepticism or opposing viewpoints in aim to clarify on fake news content. Analysis of a network can show one the difference between each class of users based on components and individuals within components.
The difference in networks for each user type stems with the overall connections and variance within the network. Friendships networks have supporting (+) edges between a set of nodes (people). Diffusion networks (or networks that propagate fake information) have positive edges within the nodes around the point node; there is no cross node interaction. Credibility networks (individuals that fall under clarifier users) fall into a cross-node, both supporting (+) and opposing (-) edges. Users can be identified via the network structure they fall under relative to the individuals they engage with on social media sites.
The third component is the temporal component of the framework. The temporal component looks at when fake news media is being published and propagated. This low level feature (time) can give us insight into understanding if there are specific network properties relative to time taken to have a piece of propaganda propagate through a network. Because many of these nodes are bots, Shu, et al suggests that we can use the temporal component to identify fake media content.
Overall, the framework proposed to identify fake media content deals with the basics of communication and credibility taught to young kids. What is the content, who is posting the content and when is the content being posted are all basic credibility questions when evaluating a source. Those same methodologies can be used as questions to identify fake media.
Source: https://arxiv.org/pdf/1804.10233.pdf
Studying Fake News via Network Analysis: Detection and Mitigation: Kai Shu, H. Russell Bernard and Huan Liu (2018)