Abstract:

Traditional Graph Neural Network (GNN) models cannot model complex relationships between nodes and exhibit weak performance in handling large-scale graphs. They are unable to effectively extract representative subgraphs from large-scale graphs, which consequently leads to a low accuracy in both training and inference processes. To address these issues, a rumor detection model called MFLAN, based on an improved Graph Attention Network (GAT), is proposed, which incorporates a feature fusion method equipped with an attention mechanism. This approach assigns different weights to each feature and performs a weighted summation operation on the original features to obtain a fused feature vector. Additionally, a positive positional encoding is introduced, which enables the model to capture positional information representations. Subsequently, learnable parameter matrices are incorporated, enabling the model to automatically learn and optimize parameter values during the training process. Finally, the attention scores are sparsified by setting the attention weights of certain unimportant nodes in the large-scale graph to zero, thereby completing the construction of the MFLAN model. Experimental results demonstrate that the proposed model achieves an accuracy of 97.71% on the Ma-Weibo dataset and 97.10% on the Weibo23 dataset. Compared to the accuracies achieved by the Dir-GNN model, these represent improvements of 1.07% and 1.12%, respectively. Compared to other rumor detection models, MFLAN exhibits superior performance across all evaluation metrics.

Key words: rumor detection, Sina Weibo, information dissemination, feature fusion, Graph Attention Network (GAT)

摘要：

传统的图神经网络(GNN)模型缺乏对节点之间复杂关系的建模能力, 且对大规模图的处理能力较弱, 无法有效地从大规模图中提取代表性的子图, 由此导致训练和推理的精确率不高。为此, 提出一种基于改进图注意力网络(GAT)的多特征融合谣言检测模型MFLAN。首先, MFLAN通过加入带有注意力机制的特征融合方法, 为每个特征赋予不同的权重, 对原始特征进行加权求和操作, 获得融合后的特征向量; 其次, 加入正值位置编码, 使模型可以获取位置信息表示; 然后, 引入可学习的参数矩阵, 使得模型在训练过程中自动地学习和优化参数值; 最后, 对注意力分数进行稀疏化, 将大规模图中部分不重要节点的注意力置为0, 以此实现MFLAN模型。实验结果表明, MFLAN模型在Ma-Weibo和Weibo23数据集上的准确率分别达到97.71%和97.10%, 相较于Dir-GNN模型分别提升1.07%和1.12%, 与其他谣言检测模型相比, MFLAN各项性能指标均有提升。

关键词: 谣言检测, 新浪微博, 信息传播, 特征融合, 图注意力网络