Spectral Clustering-Guided News Environments Perception for Fake News Detection

Fake news detection is crucial for preventing the spread of misinformation on social media. Whereas existing researches tend to ignore some of the hidden signals in the news environments, which verify the authenticity of the news posts by zooming in textual signals and incorporating external knowledge. In this study, we propose the Spectral clustering Environments and data Augmentation for Fake News Detection (SEAFND) method, which provides novelty and popularity from the news posts for fake news detection to boost the detection accuracy. We first leverage the graph theory-based spectral clustering approach to obtain the center-of-mass vectors and entropies to capture the complex associations and implicit evidences among the news on social media. Then, we introduce a shared parameter multitask learning framework that treats different news environments as independent tasks and architects GRU bootstrapping modules with attention mechanisms to help aggregate features from different environments efficiently and interpretably. Finally, we provide textual perspective and stylistic perspective approaches during the detection process and soften the loss terms in multiple environments to alleviate the strict constraints, thus making it more compatible with the fake news detection task. Compared with the state-of-the-art methods, the SEAFND method improves the detection performance by 1% to 2% on the multidomain datasets Ch-9 and En-3.