基于广度-深度采样和图卷积网络的谣言检测方法  

Rumor detection method based on breadth-depth sampling and graph convolutional networks

作　　者：王友卫[1] 王炜琦 凤丽洲 朱建明[1] 李洋[1] WANG Youwei;WANG Weiqi;FENG Lizhou;ZHU Jianming;LI Yang(School of Information,Central University of Finance and Economics,Beijing 100081,China;School of Statistics,Tianjin University of Finance and Economics,Tianjin 300222,China)

机构地区：[1]中央财经大学信息学院,北京100081 [2]天津财经大学统计学院,天津300222

出　　处：《浙江大学学报（工学版）》2024年第10期2040-2052,共13页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(61906220);国家社科基金资助项目(18CTJ008);教育部人文社科资助项目(19YJCZH178);中央财经大学新兴交叉学科建设项目;内蒙古纪检监察大数据实验室2020-2021年度开放课题资助项目(IMDBD202002,IMDBD202004).

摘　　要：现有谣言检测方法存在早期数据丢失、特征利用不充分问题,为此提出新的检测方法.为了充分挖掘事件的早期传播特征,提出广度采样方法并构建与事件对应的传播序列,利用Transformer挖掘长距离评论间的语义相关性并构建事件的传播序列特征.为了有效挖掘事件的传播结构特征,提出基于路径长度的深度采样方法,构建事件对应的信息传播子图和信息聚合子图,利用图卷积网络在挖掘图结构特征方面的优势,获得与事件对应的传播结构特征.将事件对应的传播序列特征表示与传播结构特征表示进行拼接,得到事件对应的最终特征表示.在公开数据集Weibo2016和CED上开展所提方法的有效性验证实验.结果表明,所提方法普遍优于现有典型方法.与基线方法相比,所提方法的准确率和F1值均有显著提升,所提方法在谣言检测领域的有效性得到验证.A new detection method was proposed to resolve the problems of early data loss and insufficient feature utilization in the field of rumor detection.In order to fully extract early propagation features of events,a breadth sampling method was proposed,and propagation sequences corresponding to events were constructed.A Transformer was utilized to explore semantic correlations between long-distance comments and to construct propagation sequence features for events.In order to effectively uncover the structural features of event propagation,a depth sampling method based on path length was proposed,and information propagation subgraphs and information aggregation subgraphs corresponding to events were constructed.The advantage of graph convolutional networks in exploring graph structural features was leveraged to obtain the propagation structure features corresponding to events.Feature representation of the propagation sequence and propagation structure for events were concatenated to obtain the ultimate feature representation.Validation experiments for the proposed method were conducted on two public datasets(Weibo2016 and CED).Results show that the proposed method is generally superior to existing typical methods.Compared to baseline methods,the proposed method has significant improvements in accuracy and F1 score,validating the effectiveness of the method in the field of rumor detection.

关 键 词：谣言检测 图卷积网络 广度采样 深度采样 注意力机制 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]