在社交媒体平台上部署有效的谣言检测机制，对于维护公共舆论秩序与保障社会稳定有着重要意义。前沿的面向社交媒体平台的谣言检测方法常使用深度学习技术，利用图神经网络提取谣言传播过程中的结构特征并进行分类，取得了可观的性能表现。而此类方法存在以下不足之处：（1）其中的图神经网络特征学习模型在提取图结构重要特征及建模双向图结构特征的交互上存在不足；（2）使用的输入特征单一，仅利用了事件传播过程中的文本与图结构特征，忽视了其他有效特征；（3）训练过程简单，仅使用交叉熵分类损失函数，无法充分发挥深度学习模型潜力。本文所展开的研究工作便是针对上述不足作出改进，设计出更精准、鲁棒性更高的谣言检测方法，其主要研究内容及贡献如下：（1）提出了一种基于图同构神经网络的谣言检测方法。该方法使用双向图同构神经网络以分别学习谣言传播过程图结构中的信息传播与反馈聚合特征，并使用跨方向注意力层以充分建模上述双向图结构特征间的交互。最后使用全连接层输出预测标签的概率分布。实验结果表明，该方法相较其他方法达到了更高的检测精度，在四个数据集上最大的分类准确率提升幅度为1.0%。且早期谣言检测实验结果验证了该方法检测准确率随停止时间前移下降幅度较小，说明该方法有着良好的鲁棒性。（2）在第一部分工作的基础上，提出了一种基于对比学习与多类特征融合的谣言检测方法。该方法使用核心子树特征，引入了与谣言事件中重要节点相关的用户信息、内容信息以及传播过程信息，有效地拓展了输入数据。之后，该方法对多类特征进行充分融合后使用全连接层输出预测标签的概率分布。在训练阶段，该方法使用有监督与无监督对比学习以进一步提升性能。实验结果表明，该方法在检测精度上有着进一步的提升，在数据集上相较基线模型最大的分类准确率提升幅度为2.8%。经过必要修改后，该方法的变种在早期谣言检测任务上同样有着良好的表现，有着优秀的鲁棒性。综上，本文提出的方法对现有谣言检测方法中存在的三点不足作出了针对性的改进，达到了更高的检测精度和鲁棒性。
Current social media platforms provide convenient and fast way for information exchanging for hundreds of millions of users. However, this also creates conditions for the widespread propagation of rumors. If left unchecked, the spread of rumors can endanger public opinion order and even social stability. Therefore, deploying effective rumor detection mechanisms on social media platforms to identify and prevent the spread of rumors in cyberspace is of significant importance for maintaining public opinion order and ensuring social stability.Cutting-edge rumor detection methods for social media platforms often use deep learning techniques, employing graph neural networks to extract structural features in the rumor propagation process and achieving notable performance. However, these methods have the following disadvantages: (1) The designed graph neural network feature extractors are insufficient to fully capture important features within the graph structure and model the interactions between features from two directions; (2) They only consider propagation structural features, ignoring other crucial information in rumor events; (3) The training process is crude, using only the cross-entropy classification loss function, which cannot fully exploit the potential of deep learning models. These shortages limit the performance of existing rumor detection methods and indicate possible directions for improvement.The work in this thesis aims to make improvements toward the above disadvantages and designs more accurate and robust rumor detection methods. The main research content and contributions are as follows:(1) A rumor detection method based on graph isomorphism networks is proposed. This method uses bidirectional graph isomorphism networks to respectively extract information propagation and feedback aggregation features in rumor events, and employs the cross-directional attention layer to fully model the interactions between the above two kinds of features. Finally, a fully connected layer outputs the probability distribution of the predicted labels. Experimental results show that this method achieves higher detection accuracy and robustness compared to other baseline methods, with the highest classification accuracy improvement of 1.0% across four datasets. In addition, the early rumor detection experiment shows its accuracy decreases with a small magnitude as the stopping time becomes earlier, which proves the proposed method has good robustness.(2) Building on the first part of the work, a rumor detection method based on contrastive learning and multi-feature fusion is proposed. This method uses the kernel subtree features, incorporating user information, content information, and propagation process information related to key nodes in rumor events, effectively expanding the model‘s input data. Subsequently, the low-rank multimodal fusion method is used to merge kernel subtree features, rumor source text features, and propagation structural features. Finally, a fully connected layer outputs the probability distribution of the predicted labels. During training, this method employs both supervised and unsupervised contrastive learning to further enhance performance. Experimental results show that this method achieves further improvements in detection accuracy and robustness compared to other methods, with the highest classification accuracy improvement of 2.8% across four datasets. After necessary adjustments, variants of the proposed method also achieves outstanding performance on the early rumor detection task, showing good robustness.The first part of this work makes improvements for the aforementioned first disadvantage, while the second part further tackles the second and third ones. In summary, the methods proposed in this study make improvements to address the three disadvantages present in existing rumor detection methods and achieves higher detection accuracy and robustness.
