FNDPro: Evaluating the Importance of Propagations during Fake News Spread

Herun Wan, Ningnan Wang, Xiang Zhao, Rui Li, Hui Yang, Minnan Luo

Abstract: Existing fake news detection models fall into two categories: content-based and graph-based. Content-based models identify fake news depending on news content, which may fail to determine fake news with disguised content. Graph-based models adopt extra media to construct graphs, which provide social context to identify fake news. However, existing graph-based models treat each media equally, neglecting the echo chamber phenomenon where most media have the same opinion or similar content. A model that can dynamically evaluate the contribution of each propagation and find critical ones during news spread is needed. To this end, we proposed FNDPro, which models the news propagation process as a heterogeneous dynamical graph. The key is that it models news as the first propagation and \(\ell \)-hop neighbors as the \((\ell +1)\)-th propagation. FNDPro contains a multi-modality encoder to encode each media and a propagation encoder to encode each propagation. FNDPro then employs a propagation transformer module to make every propagation embedding interact and obtain the importance score of each propagation. FNDPro achieves the best performance on three real-world datasets. Further experiments show the propagation transformer is helpful. Notably, FNDPro shows great generalization capabilities and can detect fake news even when news media are limited and manipulated. (Resources are available at https://github.com/whr000001/FNDPro.)