Lay it Out: Detecting of Fake News Publishers through Website Structure Data

Abstract

We propose a novel website structure based domain-level fake news detection model that has performance results surprisingly comparable to that of existing content-based methods. Through feature analysis, we highlight that fake news sites have more clustered subpages and more {\it ads} links, whereas traditional news sites are more substantive and more likely to contain {\it staff} links. We then illustrate that the structural model has a higher overall false positive rate compared to content-based methods, which have a higher false negative rate for domains that are more recent, more popular, and is conservative-leaning. Additionally, we also show that all model performance is dependent on the strictness in definitions of fake and traditional news sites. Specifically, model performance is higher when these definitions are more restrictive. Finally, we demonstrate that the performance of existing content-based models improve significantly by incorporating structural features, particularly when the definitions for fake and traditional news sites are lax.