Topology Imbalance and Relation Inauthenticity Aware Hierarchical Graph Attention Networks for Fake News Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 4687–4696 October 12–17, 2022. 4687 Topology imbalance and Relation inauthenticity aware Hierarchical Graph Attention Networks for Fake News Detection Li Gao1, Lingyun Song∗1, 2, Jie Liu1, Bolin Chen1, 2, Xuequn Shang∗1, 2 1School of Computer Science, Northwestern Polytechnical University, Xi’an, China 2Key Laboratory of Big Data Storage and Management, Ministry of Industry and Information Technology, Xi’an, China {lysong,shang}@nwpu.edu.com Abstract Fake news detection is a challenging problem due to its tremendous real-world political and social impacts. Recent fake news detection works focus on learning news features from News Propagation Graph (NPG). However, lit- tle attention is paid to the issues of both au- thenticity of the relationships and topology im- balance in the structure of NPG, which trick existing methods and thus lead to incorrec- t prediction results. To tackle these issues, in this paper, we propose a novel Topology imbalance and Relation inauthenticity aware Hierarchical Graph Attention Networks (TR- HGAN) to identify fake news on social me- dia. Specifically, we design a new topology imbalance smoothing strategy to measure the topology weight of each node. Besides, we adopt a hierarchical-level attention mechanis- m for graph convolutional learning, which can adaptively identify the authenticity of relation- ships by assigning appropriate weights to each of them. Experiments on real-world datasets demonstrate that TR-HGAN significantly out- performs state-of-the-art methods. 1 Introduction Fake news is a news article that is intentional- ly and verifiably false and could mislead readers (Rashkin et al., 2017; Zhou and Zafarani, 2020). The widespread of fake news can immensely af- fect cyberspace security and even social stability (Lazer et al., 2018). For example, fake news “all of Walmart’s e-commerce stores would include a pay with litecoin option beginning October 1"1 causes Litecoin’s prices surged by over 37% to $236 and Bitcoin’s prices fell below $44,000 (Kogan et al., 2021), resulting in huge losses for many investors. Various fake news detection methods are devel- oped to classify whether the news is fake or not, ∗Corresponding authors: Lingyun Song, Xuequn Shang 1https://genesisblockhk.com/fake-news-litecoin-and- walmart-are-not-partnering-in-payments/ which can be roughly divided into news content- based methods and social context-based method- s. Existing news content-based methods (Qazvini- an et al., 2011; Maddock et al., 2015; Jin et al., 2013; Wu et al., 2015; Ma et al., 2017) typically fo- cus on mining lexical and syntactic features (Feng et al., 2012; Potthast et al., 2017; Conroy et al., 2015) from news contents, ignoring the rich struc- tural information of news propagation. To address this limitation, social context-based methods (Yuan et al., 2019; Yang et al., 2021; Shu et al., 2019; Yuan et al., 2020) tend to learn the feature repre- sentations by the structure information of News Propagation Graph (NPG), where NPG consists of multiple types of nodes (e.g., news, comments and users) and relationships (e.g., follower, retweet and friendship). (a) (b) (c) News2 … … … User1 User1 User3 User4 User3 User2 User1 News2 News3 … News1 Comment1 Figure 1: Three types of unauthentic relation …

Content truncated to 500 words