Abstract: To address the existing graph neural network models?? limitations in effectively modeling complex multi-node relationships during rumor propagation and insufficient exploration of multi-modal feature information within nodes for multi-modal fake news detection, this paper proposes a hypergraph-based rumor detection model incorporating multi-modal node embeddings. Firstly, the method fuses multi-modal features, including text, image, and sentiment to obtain more informative initial node embeddings. Then, it feeds the fused node representations into the constructed multi-relational hypergraph to capture group-level interaction patterns beyond pairwise relations. Finally, it introduces a gated dual-attention mechanism that adaptively assigns weights to key nodes and hyperedges and highlights high-importance factors. The resulting high-level node representations are used as classifier inputs, ultimately improving the identification of news relationships and propagation patterns. Experimental evaluations demonstrate that the proposed method achieves significant performance improvements on three publicly available datasets, obtaining accuracies of 94.46%, 97.36% and 93.86%, respectively. Furthermore, the method shows promising effectiveness in early rumor detection tasks, highlighting its advantages in multi-modal information fusion and complex relationship modeling.

Key words: graph neural networks, rumor detection, multi-modal fusion, attention mechanism

摘要： 针对现有图神经网络模型在多模态新闻谣言检测任务中难以有效建模谣言传播过程中的复杂多节点关系、且对节点自身所含的多模态特征信息挖掘不足的问题，提出一种融合图文及情感等多模态特征的节点嵌入的超图网络谣言检测模型。该方法融合包含图文以及情感等多模态特征，得到信息更充分的初始节点嵌入；将融合后的节点特征导入构建的多重关系的超图结构中，刻画超过成对关系的群体交互模式；引入带门控机制双重注意力机制，自适应给关键节点和超边分配权重并突出高权重要素；得到的节点高层表示作为分类输入，最终提升新闻关系和传播模式的识别能力。通过实验验证，所提出的方法在三个公开数据集上取得了显著的性能提升，准确率分别为94.46%、97.36%和93.86%。此外，该方法在早期谣言检测任务中也表现出一定的有效性，展示了其在多模态信息融合及复杂关系建模方面的优势。

关键词: 图神经网络, 谣言检测, 多模态融合, 注意力机制