A unified cross-source context enhancement model for multi-source fake news detection

Abstract

The growing prevalence of misinformation across digital platforms poses a serious threat to public information integrity. While existing fake news detection methods have predominantly focused on single-source data, such approaches often fail to address the complexities introduced by multi-source information originating from diverse platforms. Detecting fake news in this multi-source context remains a relatively underexplored challenge. Traditional cross-domain methods typically transfer domain-invariant features using single-source modeling strategies, yet they neglect the rich contextual cues available across sources—an aspect critical for enhancing detection accuracy. To overcome these limitations, we propose a unified framework that effectively captures global contextual information by jointly leveraging intra-source and inter-source interactions. Our model integrates two principal components: (1) a cross-source global context learning module employing a context-augmented transformer to model long-range dependencies among multi-source instances, and (2) a dual-level contrastive learning mechanism that aligns representations at both local and global levels, reducing inconsistencies across feature spaces and source domains. Extensive experiments conducted on publicly available multi-source datasets demonstrate that our method achieves substantial improvements over existing state-of-the-art approaches. Specifically, it yields an approximate 5% gain in classification accuracy compared to leading models such as LIMFA, highlighting the robustness and effectiveness of our framework in advancing multi-source fake news detection.