Scholarlink

<p>With the rapid evolution of social media platforms, the proliferation of fake news in various forms increasingly misleads readers. Existing multimodal and graph-based methods have successfully leveraged either cross-modal interactions or propagation networks to identify fake news. However, they struggle to generalize real-world situations, where news posts are rich in multiple modalities and widely disseminated across social networks. Furthermore, they overlooked the complementarity between multimodal information and propagation networks, as detection clues in multimodal news often emerge within comment propagation chains. To address these issues, we propose the Multimodal Uncertainty Graph Contrastive Learning (MUGCL) framework, the first attempt to integrate textual, visual, and propagation networks into a unified contrastive learning approach for fake news detection. Specifically, MUGCL constructs two parallel propagation chains, rooted in the textual and visual modalities of news, to capture propagation patterns of detection clues. Furthermore, we introduce an uncertainty-aware graph contrastive learning strategy to model collaborative interactions between different propagation chains and adaptively aggregate cross-modal uncertainties. Extensive experiments demonstrate that MUGCL significantly enhances fake news detection performance. Additionally, studies confirm that MUGCL effectively captures diverse real-world feedback to distinguish multimodal news, reduces dependence on news content, and improves generalization to complex social contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Towards real-world multimodal propagation networks: Multimodal uncertainty graph contrastive learning for fake news detection

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Siyan Nie,
  • Zhi Zeng

摘要

With the rapid evolution of social media platforms, the proliferation of fake news in various forms increasingly misleads readers. Existing multimodal and graph-based methods have successfully leveraged either cross-modal interactions or propagation networks to identify fake news. However, they struggle to generalize real-world situations, where news posts are rich in multiple modalities and widely disseminated across social networks. Furthermore, they overlooked the complementarity between multimodal information and propagation networks, as detection clues in multimodal news often emerge within comment propagation chains. To address these issues, we propose the Multimodal Uncertainty Graph Contrastive Learning (MUGCL) framework, the first attempt to integrate textual, visual, and propagation networks into a unified contrastive learning approach for fake news detection. Specifically, MUGCL constructs two parallel propagation chains, rooted in the textual and visual modalities of news, to capture propagation patterns of detection clues. Furthermore, we introduce an uncertainty-aware graph contrastive learning strategy to model collaborative interactions between different propagation chains and adaptively aggregate cross-modal uncertainties. Extensive experiments demonstrate that MUGCL significantly enhances fake news detection performance. Additionally, studies confirm that MUGCL effectively captures diverse real-world feedback to distinguish multimodal news, reduces dependence on news content, and improves generalization to complex social contexts.