Implementation and Evaluation of an Interpretable Fake News Detector

Abstract

Interpretability is an important element of fake news detection so that readers can assess the credibility of news by themselves. We implemented a naive Bayes fake news detection model proposed by Granik and Mesyur and evaluated it with the LIAR dataset in terms of recall, effect of stop words, and interpretability. The recall was affected by the imbalanced data and eliminating stop words did not improve the accuracy but slightly deteriorated it. Some high probability words were interpretable as reasons for fake news but longer phrases had better be considered as clues for fake news.