一种基于动态异构图的谣言检测模型  

Rumor Detection Model Based on Dynamic Heterogeneous Graph Neural Network

作　　者：朱文龙[1,2] 陈羽中 饶孟宇[2] ZHU Wenlong;CHEN Yuzhong;RAO Mengyu(College of Computer and Data Sciences,Fuzhou University,Fuzhou 350116,China;Fujian Provincial Key Laboratory of Network Computing and Intelligent Information Processing,Fuzhou 350116,China)

机构地区：[1]福州大学计算机与大数据学院,福州350116 [2]福建省网络计算与智能信息处理重点实验室,福州350116

出　　处：《小型微型计算机系统》2024年第2期319-326,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61672158,61972097,U21A20472)资助;福建省科技重大专项专题项目(科教联合)(2021HZ022007)资助;福建省高校产学研合作项目(2021H6022)资助;福建省自然科学基金项目(2020J01494)资助。

摘　　要：随着互联网技术和自媒体行业的快速发展,人们可以方便快捷地从社交媒体中获取最新信息,但也让更多的谣言在网络中盛行.现有谣言检测模型多从文本内容、用户信息和传播模式中挖掘有效特征.然而,现有模型未充分学习文本的语义信息和谣言传播过程中的结构信息,并忽略了谣言传播的动态过程.针对上述问题,本文提出一种基于动态异构图的谣言检测模型DHGNN(Dynamic Heterogeneous Graph Neural Network).首先,为了增强帖子的文本语义表示,本文提出一种多级注意力网络,引导模型关注源帖子和相应评论中关键的词和句子,充分学习源帖与相应评论之间的语义关联.其次,引入了基于异构图的图神经网络,通过对异构传播图中的用户、帖子节点和转发(或评论)关系进行建模,为不同类型的节点和边生成特定的表示,充分学习异构传播图中的结构信息.最后,提出一种基于旋转记忆单元的时序注意力,分别为每个异构传播图快照建立记忆,捕获谣言动态传播的演化模式.在Twitter15、Twitter16数据集上的实验结果表明,DHGNN模型的性能优于最新的对比模型.With the rapid development of Internet technology and the self-media industry,people can easily and quickly obtain the latest information from social media,but it also allows more rumors to prevail on the Internet.Most existing rumor detection models focused on mining effective features from textual content,user profiles,and propagation patterns.However,these methods do not fully learn the semantic information of the text and structural information in the process of rumor propagation.They also ignore the dynamic process of rumor propagation.To address these issues,this paper proposes a rumor detection model based on dynamic heterogeneous graph neural network(DHGNN).First,we propose a multi-level attention network to enhance the semantic representation of posts.The multi-level attention network can guide DHGNN to pay more attention to key words and sentences in the source posts and comments.Therefore,DHGNN can fully learn the semantic relationships between the source posts and the corresponding comments.Second,we propose a heterogeneous graph neural network by modeling users,post nodes,and retweet(or comment)relationships in a heterogeneous propagation graph.DHGNN can generate specific representations for different types of nodes and edges and fully learn structural information in heterogeneous propagation graphs.Third,we propose a temporal attention mechanism based on the rotating memory unit,which can establish memory for each heterogeneous propagation graph snapshot and capture the dynamic evolution pattern of rumor propagation.Experimental results on Twitter15,Twitter16 datasets show that DHGNN outperforms other state-of-the-art models.

关 键 词：谣言检测 多级注意力 异构传播图 图神经网络 时序注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]