DHCF: Dual disentangled-view hierarchical contrastive learning for fake news detection on social media

Widespread fake news on social media threatens public security and the cyber environment, making fake news detection an essential area of study. The majority of existing fake news detection methods rely on news content (e.g., text and images) and/or social contexts (e.g., comment interactions between posts) to determine the veracity of news. However, existing methods still have the following drawbacks: (1) Overreliance on sufficient reliable labeled data. (2) Lack of robustness to noise and fraudster-designed harmful disguises. (3) Inability to differentiate between the multiple intentions behind retweet and comment behaviors, resulting in generating entangled representations. To address the above aforementioned three issues, we introduce contrastive learning and disentangled representation learning for fake news detection. Specifically, to mine supervised signals from unlabeled data and improve the model's robustness, we design a hierarchical contrastive learning framework that includes multiple data augmentation strategies and three contrastive learning tasks. In addition, to infer the latent intentions of retweets and comments between posts, we propose the disentangled graph encoder (Disen-GraphEnc) and disentangled sequence encoder (Disen-SeqEnc). Extensive experiments demonstrate the superiority of our model over other state-of-the-art methods and is resistant to limited training data and noise attacks. Our code is available on the GitHub (https://github.com/senllh/DHCF).