基于多特征图与GCN-GAT的谣言检测模型

Rumor Detection Model Based on Multi-Feature Graphs and GCN-GAT

摘要: 社交媒体中的谣言文本通常是非正式且语法不连贯的，这为准确提取语义信息带来了巨大挑战。为解决这一问题，本文提出了RoBERTa-MGAT模型，一种融合多特征图构建、并行图神经网络编码器与基于注意力的融合机制的谣言检测模型。具体而言，该模型利用RoBERTa和Word2Vec生成丰富的词向量表示，并通过构建三种异构图(词性图、词共现图和语义依存图)从不同角度捕捉多样化的语言特征。针对每个图结构，本文并行采用图卷积网络(Graph Convolutional Network, GCN)和图注意力网络(Graph Attention Network, GAT)来联合学习互补的结构化表征，GCN专注于捕获局部平滑特征，而GAT则用于建模特征异构性与重要性。最后通过自注意力机制聚合所有图的输出表示，使模型能够有效整合多视角特征。在Weibo20和Weibo21两个公开谣言检测数据集上的实验结果表明，RoBERTa-MGAT在准确率和F1分数上均优于同类型模型，展现出卓越的性能。

Abstract: Rumor texts on social networks are often informal and grammatically incoherent, posing significant challenges for extracting accurate semantic information. To address this issue, we propose RoBERTa-MGAT, a rumor detection model that integrates multi-feature graph construction with parallel graph neural encoders and an attention-based fusion mechanism. Specifically, the model leverages RoBERTa and Word2Vec to generate rich word embeddings and constructs three heterogeneous graphs—a part-of-speech graph, a word co-occurrence graph, and a semantic dependency graph—to capture diverse linguistic features from different perspectives. For each graph, Graph Convolutional Network (GCN) and Graph Attention Network (GAT) are applied in parallel to jointly learn complementary structural representations, with GCN capturing local smoothness and GAT modeling feature heterogeneity and importance. The outputs from all graphs are then aggregated using a self-attention mechanism, allowing the model to effectively integrate multi-view features. Experimental results on two public rumor detection datasets, Weibo20 and Weibo21, demonstrate that RoBERTa-MGAT achieves superior performance, outperforming existing state-of-the-art methods in both accuracy and F1-score.