融合元路径学习和胶囊网络的社交媒体谣言检测方法

Rumor Detection Based on Meta-Path Learning and Capsule Network

  • 摘要: 以源推特文本为研究对象,深度挖掘推特正文内容的语义信息,并强调谣言在具有异质性的社交网络传播过程中存在的结构特征,以达到提升谣言检测效果的目的。采取基于One-Hot Encoding的词嵌入方法,结合Multi-head attention机制实现推特正文内容初级语义特征的提取,并进一步基于胶囊网络(CapsNet)构建内容胶囊(content-capsule)模块实现对正文内容深度语义特征的提取,结合图卷积胶囊(GCN-Capsule)模块实现谣言在社交网络中传播结构特征的提取,将两种胶囊向量采用一种动态路由机制进行融合,进一步丰富输入特征,之后输出源推特的分类结果,进而实现源推特的谣言检测。实验结果显示,该模型对谣言识别的正确率达93.6%。

     

    Abstract: Aiming at taking the source Twitter texts as the research object, this paper deeply explores semantic information of Twitter body content and emphasizes structural features of heterogeneous rumor spreading social networks, so as to improve rumor detection effect. This paper combines one-hot encoding word embedding method and multi-head attention mechanism to extract primary semantic feature of source Twitter text content. Furthermore, the content-capsule module is constructed based on CapsNet to extract the deeper semantic features of text content, and the structure features of rumor propagation in social networks are extracted by combining with GCN-capsule module. In order to further enrich the input, two kinds of capsule vectors are fused with a dynamic routing mechanism. And then the classification results of source tweets are output,and source tweets rumors detection is finished. Experimental results show that the accuracy of the model proposed in this paper reaches 93.6%.

     

/

返回文章
返回