Multi-depth Graph Convolutional Networks for Fake News Detection

Fake news arouses great concern owing to its political and social impacts in recent years. One of the significant challenges of fake news detection is to automatically identify fake news based on limited information. Existing works show that only considering news content and its linguistic features cannot achieve satisfactory performance when the news is short. To improve detection performance with limited information, we focus on incorporating the similarity of news to discriminate different degrees of fakeness. Specifically, we propose a multi-depth graph convolutional networks framework (M-GCN) to (1) acquire the representation of each news node via graph embedding; and (2) use multi-depth GCN blocks to capture multi-scale information of neighbours and combine them by attention mechanism. Experiment results on one of the largest real-world public fake news dataset LIAR demonstrate that the proposed M-GCN outperforms the latest five methods.