| 引用本文： |
-
卫玲蔚,胡斗,鲍祎楠,周薇,杨近朱,虎嵩林.基于时序特征和结构特征的社交网络谣言检测方法[J].信息安全学报,2026,11(2):289-299 [点击复制]
- WEI Lingwei,HU Dou,BAO Yinan,ZHOU Wei,YANG Jinzhu,HU Songlin.Jointly Exploiting Temporal and Structural Features for Rumor Detection on Social Media[J].Journal of Cyber Security,2026,11(2):289-299 [点击复制]
|
|
| |
|
|
| 本文已被：浏览 8次 下载 4次 |
码上扫一扫！ |
| 基于时序特征和结构特征的社交网络谣言检测方法 |
|
卫玲蔚1,2, 胡斗1,3,2, 鲍祎楠1,2, 周薇1, 杨近朱1,2, 虎嵩林1,2
|
|
|
| (1.中国科学院信息工程研究所 北京 中国 100085;2.中国科学院大学网络空间安全学院 北京 中国 100049;3.华北计算机系统工程研究所 北京 中国 100083) |
|
| 摘要: |
| 随着社交网络的快速发展,越来越多的人在社交网络平台获取或分享信息。但是,在收获便利的同时,也为谣言提供了新的传播媒介。谣言的传播严重影响网络空间清朗环境的建设,开展自动化谣言检测至关重要。现有基于深度学习的谣言检测模型往往基于内容特征或传播特征展开,而这些基于传播特征的模型要么只关注传播过程中的时序关系,或是仅挖掘谣言传播网络的结构特征来识别谣言,不能很好地学习一个全面的特征表示描述谣言传播过程中的时间和空间变化,限制了谣言检测的性能。针对此问题,本文提出一种通用的基于时序特征和结构特征的谣言检测方法,共同探索谣言传播过程中的时间模式与传播树结构特征,学习一个全面的谣言特征表示,提高谣言检测的性能。为评估模型的有效性,本文在3个谣言真实数据集进行实验。实验结果表明,本文方法平均获得了4.8%准确率绝对提升。大量实验验证了所提方法在谣言检测的有效性。 |
| 关键词: 谣言检测 时序特征 传播结构学习 门控循环单元 图卷积神经网络 社交网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2026.03.18 |
| 投稿时间：2020-12-25修订日期：2021-03-04 |
| 基金项目:本课题得到国家自然科学基金(No.62102412)和国家重点研发计划(No.2022YFC3302102)资助。 |
|
| Jointly Exploiting Temporal and Structural Features for Rumor Detection on Social Media |
|
WEI Lingwei1,2, HU Dou1,3,2, BAO Yinan1,2, ZHOU Wei1, YANG Jinzhu1,2, HU Songlin1,2
|
| (1.Institute of Information Engineering, Chinese Academy of Sciences, Beijing 100085, China;2.School of Cyber Security, University of Chinese Academy of Sciences, Beijing 100049, China;3.National Computer System Engineering Research Institute of China, Beijing 100083, China) |
| Abstract: |
| With the rapids development of the social network, more and more people obtain or share information on social network platforms. Unfortunately, the convenient environment of social network platforms has also provided a new medium for the spread of rumors. The spread of rumor has become a significant challenge that seriously undermines the credibility of information in social network and posts a threat to building a clear cyberspace environment. Automatic rumor detection is essential for timely prevention of rumor spread and maintaining social stability. The existing deep learning-based rumor detection models have been developed based on content characteristics or propagation characteristics including temporal features and structural features. However, most of these detection models either only model the temporal information in rumor propagation or only focus on the structure features of rumor propagation to identify rumors. This limitation cannot learn a comprehensive eigenvector representation well and hinders the performance of rumor detection. To alleviate the above problem, in this paper, we propose a novel graph-based rumor detection model. It combines the power of graph networks and sequence models to jointly model both structural features and temporal patterns in rumor propagation. Specifically, based on the textual features extracted by embedding layer and propagation information, we utilize a time-aware bidirectional gated recurrent unit to explore temporal features and a graph convolutional network to learn structural features. Then, we combine them to make prediction. By doing so, the model can learn a comprehensive representation of rumor characteristics, enabling it to detect rumors with greater accuracy. In addition, the model can effectively alleviate the time mode distortion caused by pruning. To evaluate the performance of the proposed model, we conduct experiments on three real-world rumor detection benchmark datasets. The experimental results show that the proposed method achieves 4.8% average absolute improvements in terms of the accuracy score across all three datasets. Extensive experiments demonstrate the effectiveness of the proposed model for rumor detection. |
| Key words: rumor detection temporal information propagation structure learning gated recurrent unit graph convolutional network social network |
|
|
|
|
|
