基于图卷积网络的多特征融合谣言检测方法

Multi-feature Fusion Rumor Detection Method Based on Graph Convolutional Network

目前,大部分谣言检测工作主要基于Twitter或新浪微博原文本内容、传播结构和传播文本内容进行谣言检测,忽略了原文本特征与其他特征的有效融合,以及传播用户在谣言传播过程中的作用。针对以上问题,提出了一种基于图卷积网络的多特征融合模型GCNs-BERT,模型同时融合了原文本特征、传播用户特征和传播结构特征。首先,基于传播结构和传播用户构建传播图,将多个用户属性的组合作为传播节点特征;其次,利用多个图卷积网络学习在不同用户属性组合的情况下传播图的表达,同时采用BERT模型学习原文本内容特征表达,最终与图卷积网络学习的特征相融合用于检测谣言。利用公开的新浪微博数据集进行的大量实验表明:GCNs-BERT模型明显优于基线方法。此外,在新冠疫情数据集上进行GCNs-BERT模型泛化能力实验,此数据集训练样本大小仅有新浪微博数据集的1/5,仍然取得了92.5%的准确率,证明模型具有较好的泛化能力。

At present,most rumor detection work mainly based on the original text content,communication struc-ture and communication text content of Twitter or Weibo.However,these methods ignored the effective integration of original text features with other features,as well as the role of propagating users in the process of rumor propaga-tion.Aiming at the shortcomings of the existing work,a multi-feature fusion model GCNs-BERT based on graph convolutional network was proposed,which combined the features of the original text,the propagating user and the propagating structure.Firstly,a propagation graph was constructed based on the propagation structure and the prop-agation users,and the combination of multiple user attributes was used as the propagation node feature.Then,mul-tiple graph convolutional networks were used to learn the representation of the propagation graph with different user attribute combinations,and BERT model was used to learn the feature representation of the original text content.Finally,the fusion with the features learned by the graph convolutional network was used to detect rumors.A large number of experiments using publicly available Weibo data sets showed that the GCNs-BERT model was significant-ly better than the baseline method.In addition,the generalization ability experiment of GCNs-BERT model was conducted on the novel coronavirus epidemic data set.The training sample size of this data set was only 1/5 of that of the public Weibo data set,and the accuracy rate was still 92.5%,which proved that the model had good gener-alization ability.