基于图自监督对比学习的社交媒体谣言检测  

Rumor detection on social media based on graph contrastive self-supervised learning

作　　者：乔禹涵 贾彩燕[1,2] Qiao Yuhan;Jia Caiyan(School of Computer and Information Technology,Beijing Jiaotong University,Beijing,100044,China;Beijing Key Laboratory of Traffic Data Analysis and Mining,Beijing Jiaotong University,Beijing,100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]交通数据分析与挖掘北京市重点实验室,北京交通大学,北京100044

出　　处：《南京大学学报（自然科学版）》2023年第5期823-832,共10页Journal of Nanjing University（Natural Science）

基　　金：中央高校基本科研业务费(2019JBZ110)。

摘　　要：网络社交媒体的快速发展提供了便捷的信息获取方式,但也滋生了谣言和虚假新闻,现有的谣言检测模型在有标注数据充足时能有效解决分类问题,然而谣言可用的标注数据有限,各种针对谣言特点精心设计的模型倾向于过拟合,同时,现有模型的鲁棒性不足,谣言传播者恶意破坏谣言传播结构会使模型出现分类错误.针对以上问题,采用自监督的图对比学习方法,对原始谣言传播图进行不同方式的数据增强来模拟对原图的扰动,建立自监督对比学习任务,使图编码器捕获谣言更趋本质的特征,缓解了过拟合,提高了模型的鲁棒性与泛化性能.在来源于主流社交媒体平台的三个公开数据集Twitter15,Twitter16和PHEME上进行了对比实验,实验结果显示,提出的模型的准确率比基准模型分别提高3.4%,1.8%和1.2%,证实了图自监督对比学习方法在谣言检测任务上的有效性.The rapid development of social media provides a convenient way to obtain information,meanwhile it helps the spread of rumors.Generally,with enough labeled data,existing rumor detection models can effectively solve rumor classification problems.However,due to limited labeled data of rumors,previous methods carefully designed for the characteristics of rumors tend to over-fit.Besides,existing rumor detection models are not robust enough.To solve the above problems,the graph contrastive self-supervised learning approach is adopted.A contrastive loss is defined to make graph encoders capture more essential and intrinsic features of rumors,alleviating the over-fitting and improving the robustness and generalization of the model.Experiments on three public datasets Twiter15,Twitter16 and PHEME has enhanced the accuracy of 3.4%,1.8% and 1.2% respectively compared with the baseline,confirming the effectiveness of the proposed method.

关 键 词：谣言检测 自监督学习 对比学习 图表示学习 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]