A Deep Learning Model for Early Detection of Fake News on Social Media*

Abstract

Fake news detection has recently become an important topic of research. This is due to the impact of fake news on the internet especially on social media. Numerous of the models proposed in the previous studies are based on supervised learning. Therefore, these models are unable to deal with the huge amount of unlabeled data about fake news. Few studies focused on early detection. In this study, we built a semi-supervised learning model to detect fake news on social media at an early stage. By using a semi-supervised learning, we make our model able to deal with the huge amount of unlabeled data on social media. We first built a model to extract users' opinion expressed in comments, then we used CredRank Algorithm to evaluate users' credibility and built a small network of users involved in the spread of a given news. The outputs of these three steps serve as inputs of our news classifier SSLNews. SSLNews is composed of three networks: a shared CNN, an unsupervised CNN and a supervised CNN. We used real world datasets to evaluate our model, Politifact and Gossipcop. When using 25% of labeled data, SSLNews reaches an accuracy of 72.25% on Politifact and 70.35% on Gossipcop. When using data produced in the first 10 minutes of the beginning of the spread of the news, SSLNews reaches an accuracy of 71.10% on Politifact and 68.07% on Gossipcop.