A Discriminative Graph Neural Network for Fake News Detection

Fake news detection aims to distinguish disinformation from social media. The existing fake News detection model, namely Factual News Graph (FANG), is a novel social context representation and learning framework for fake news detection. Focusing on representation learning, this inductive model is superior to conventional contextual models for its efficiency, scalability, and robustness. However, we find that the framework does not take inner/inter variants into consideration. Specifically, this method treats each news item as an independent individual and ignores the latent relationship among them, which may inevitably limit the accuracy. In order to improve the discriminative power, we propose a new model called Discriminative-FANG, which enhances the discriminative power of the model on fake news detection. By means of adding a discriminative regularization term, our model finds deep feature centers of every class and minimizes the distances between all the features and their class centers simultaneously. Thus, it leads to better learning inter-class dispersion as well as intra-class compactness. Moreover, experiments are carried out on several popular Fake News Detection datasets, which verify the better performance of our model than other popular baselines by a large margin.