融合动态传播和社区结构的社交媒体谣言检测模型  

Rumor Detection Model Based on Dynamic Propagation and Community Structure

作　　者：强子珊 顾益军[1] QIANG Zishan;GU Yijun(College of Information and Cyber Security,People’s Public Security University of China,Beijing 100032,China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100032

出　　处：《计算机工程与应用》2024年第18期198-207,共10页Computer Engineering and Applications

基　　金：中国人民公安大学网络空间安全执法技术双一流创新研究专项(2023SYL07)。

摘　　要：为解决现有谣言检测模型对时间信息利用不充分的问题,同时验证利用谣言传播的社区结构特征可以提高谣言检测模型的性能,提出一种融合动态传播和社区结构的社交媒体谣言检测模型Dy_PCRD(rumor detection model based on dynamic propagation and community structure),一方面使用图卷积网络提取谣言传播的结构特征,另一方面先根据谣言内容和传播结构划分话题社区,再使用一种新型的注意力计算方法提取谣言的社区结构特征,将二者分别输入时间注意力单元对其动态变化规律进行建模,最后基于所获得的嵌入表示对谣言进行分类。三个公开数据集上的实验结果表明,在相同条件下,相较于基线模型,其准确率及其他各评价指标均有所提升,验证了社区结构特征、动态性特征以及相关注意力计算方法对提升谣言检测模型性能的有效性。To address the insufficient utilization of time information,and to verify that the community structure features of rumor propagation can improve the performance of rumor detection models.Dy_PCRD(rumor detection model based on dynamic propagation and community structure)model is proposed which integrates dynamic propagation and commu-nity structure.On the one hand,GCN can extract structural features of rumor propagation,and on the other hand,topic communities are divided based on rumor content and propagation structure,and then a new attention calculation method is used to extract community structural features.They are inputted into temporal attention units to model their dynamic changes and classify the rumors.The experimental results on three public datasets show that under the same condition,its accuracy and other evaluation indicators have been improved compared to the baseline model,verifying the effectiveness of community structure features,dynamic features,and related attention calculation methods in improving the perfor-mance of rumor detection models.

关 键 词：谣言检测 动态图 社区结构 传播 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]