Uncertainty-aware Propagation Structure Reconstruction for Fake News Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 2759–2768 October 12–17, 2022. 2759 Uncertainty-aware Propagation Structure Reconstruction for Fake News Detection Lingwei Wei1,2 and Dou Hu1,2 and Wei Zhou1∗ and Songlin Hu1,2 1 Institute of Information Engineering, Chinese Academy of Sciences 2 School of Cyber Security, University of Chinese Academy of Sciences {weilingwei, hudou, zhouwei, husonglin}@iie.ac.cn Abstract The widespread of fake news has detrimental societal effects. Recent works model informa- tion propagation as graph structure and aggre- gate structural features from user interactions for fake news detection. However, they usually neglect a broader propagation uncertainty issue, caused by some missing and unreliable interac- tions during actual spreading, and suffer from learning accurate and diverse structural prop- erties. In this paper, we propose a novel dual graph-based model, Uncertainty-aware Prop- agation Structure Reconstruction (UPSR) for improving fake news detection. Specifically, after the original propagation modeling, we in- troduce propagation structure reconstruction to fully explore latent interactions in the ac- tual propagation. We design a novel Gaussian Propagation Estimation to refine the original deterministic node representation by multiple Gaussian distributions and arise latent interac- tions with KL divergence between distributions in a multi-facet manner. Extensive experiments on two real-world datasets demonstrate the ef- fectiveness and superiority of our model. 1 Introduction Nowadays, fake news1 has posed detrimental ef- fects on individuals and society. For example, telecommunication towers were burned due to a conspiracy theory linking COVID-19 with 5G tech- nology (Ahmed et al., 2020). To help mitigate the negative effects caused by fake news, it’s critical to develop automatic methods to detect fake news. Existing works generally leverage the user inter- actions (e.g., retweet) and shared content in a so- cial media conversation thread to detect fake news. The key principle behind such work is that users on social media share opinions, conjectures and evidence for checking fake news. Some studies * Corresponding author. 1We adopts a broad definition, i.e., fake news is false news where news broadly includes claims, statements, posts, among other types of information (Zhou and Zafarani, 2020). (Ruchansky et al., 2017; Ma et al., 2016) flatten the conversation in a chronological order to catch lin- guistic and temporal features from the propagation sequence, which does not make better use of the network properties. Some works (Ma et al., 2018; Kumar and Carley, 2019; Khoo et al., 2020; Ma and Gao, 2020) build the conversation thread with a tree structure to capture the structural patterns from the interactions of information propagation. Driven by the success of graph neural networks (Kipf and Welling, 2017), recent methods (Bian et al., 2020; Hu et al., 2021; Lin et al., 2021) regard the conversation thread as a graph structure and aggregate informative neighbors to learn a good representation for detection. However, most methods usually assume that the propagation structure is deterministic and complete at some point. In the real world, it is often the case that each sample describes a partial propagation structure that includes some missing and unreli- able interactions …

Content truncated to 500 words