An Interpretable Approach to Fake News Detection

Misinformation has long been a tool for political influence, but it has taken a new form in the information age: fake news. After exploding into public consciousness during the 2016 United States presidential election, fake news has become a reality of political life around the world, featuring heavily in the 2017 German election and the 2018 Brazilian election. Fake news poses a significant threat to civic society, and is too easily produced and quickly disseminated to be resolved by manual fact-checking. As such, fake news detection has received significant attention by machine learning and natural language processing researchers in the last years. Previous work in this field has overly relied on deep learning approaches suffering from the black-box problem, rendering them unable to articulate precisely what properties separate fake news from real news. This paper contributes to the limited work on interpretable fake news detection by engineering text-based features, applying statistical tests, and fitting and interpreting logistic regression models. The results of this paper support previous findings that fake and real news are best differentiated by metrics capturing complexity and style, that fake headlines communicate far more than real ones, and that text-based approaches can effectively discern between real and fake news