Multi-View Tri-Alignment Multi-Expert for Multi-Domain Fake News Detection

The widespread dissemination of fake news is increasingly threatening individuals and society in various domains, covering politics, health, and other aspects. It is necessary to enhance the performance of multi-domain fake news detection in the real world. Among the current works on this task, some focus on the domain difference simply, while others only consider invariant features. However, those approaches do not consider the fusion between domain differences and invariant features, imbalance under diverse cross-domain with long-tailed distribution. In this paper, Multi-view Triple-alignment Multi-expert (MTM) is proposed to address those two problems. Specifically, through extracting various representations from different views, the Multi-View Information Fusion module is applied to project the representation of a news post to the fusion feature space. Then, before the inference process, a multiple expert block is designed to discriminate on different domains for avoiding cross-domain noise. During the model training stage, through triple alignment, we align the multi-view features of each sample, the core features used to infer, and the domain features of each expert, respectively. Extensive experiments on English and Chinese datasets exhibit the effectiveness of this model. In its optimal scenario, MTM demonstrates the capability to surpass the best baseline by 5.3% in F1-socre, under conditions of rich multi-domain information.