LLM-assisted fake news detection with adaptive boosting framework incorporating contrastive learning

With the rapid development of digital technologies, the widespread spread of fake news on social media platforms may threaten social stability. Existing methods often use small models as feature extractors for news content, relying on the inherent pattern differences between true and false news as the basis for fake news detection. However, smaller models cannot perceive latent background information in news content. Additionally, fake news is much less prevalent than true news on real social platforms, which may lead to issues with data collection imbalance of true and false labels. These challenges may limit the effectiveness of existing detection methods. Therefore, we propose a Large Language Model-assisted Fake News Detection method with Adaptive Boosting, denoted as LFND-AB. Specifically, we first employ the large language model to generate analyses of the truthfulness and falseness of news text, which contains background information about the news. We then design an explainable base learner for fake news detection that utilizes contrastive learning to align the similarity between the news text and the corresponding generated analyses to obtain higher-quality fused representations. Furthermore, we design an adaptive ensemble boosting framework based on the base learner. This framework adjusts the weights of misclassified samples to improve the detection performance in the next base learner, thereby optimizing the model’s ability to handle label imbalance. The comprehensive experiments conducted on benchmark datasets demonstrate that both the proposed base learner and LFND-AB outperform State-Of-The-Art (SOTA) methods. The implementation code is available at https://github.com/cgao-comp/FEND.