基于信任门控异构图学习的社交媒体谣言检测方法

Social Media Rumor Detection Method Based on Trust-Gated Heterogeneous Graph Learning

摘要: 社交媒体谣言检测在很大程度上依赖于结构特征。然而，现有图方法在评估不同节点与交互关系的可靠性方面仍存在不足。为解决这一问题，本文提出了一种自适应信任评估框架(Adaptive Trust Evaluation Framework, ATEF)。ATEF将新闻事件建模为帖子级异构图，能够同时刻画扩散关系、反馈关系以及伪时间关系，从而充分保留传播结构信息与时间顺序信息。进一步地，本文引入了一种信任门控异构消息传递机制(Trust-Gated Heterogeneous Message Passing)，用于自适应调节不同节点及不同关系类型在信息传播过程中的贡献。通过该机制，ATEF能够增强关键传播信号，同时抑制噪声信息的干扰。实验结果表明，ATEF具有优异的检测性能。在测试集上，该模型的Accuracy和Macro-F1均达到0.9512，Fake Recall达到0.9401，显著优于BiGCN基线模型。最后，协同攻击实验与消融实验进一步验证了ATEF在复杂且高度扰动环境下具有较强的鲁棒性。

Abstract: Social media rumor detection largely relies on structural features. However, existing graph methods still have limitations in evaluating the reliability of different nodes and interaction relationships. To address this issue, this paper proposes an Adaptive Trust Evaluation Framework (ATEF). ATEF models news events as post-level heterogeneous graphs, which can simultaneously characterize diffusion relationships, feedback relationships, and pseudo-temporal relationships, thereby fully preserving propagation structure information and temporal order information. Furthermore, this paper introduces a Trust-Gated Heterogeneous Message Passing mechanism to adaptively adjust the contributions of different nodes and relationship types in the information propagation process. Through this mechanism, ATEF can enhance key propagation signals while suppressing the interference of noise information. Experimental results show that ATEF has excellent detection performance. On the test set, the model achieves an Accuracy and Macro-F1 of 0.9512, and a Fake Recall of 0.9401, which significantly outperforms the BiGCN baseline model. Finally, collaborative attack experiments and ablation experiments further verify that ATEF has strong robustness in complex and highly perturbed environments.