DDGCN: Dual Dynamic Graph Convolutional Networks for Rumor Detection on Social Media
Mengzhu Sun1, Xi Zhang1*, Jiaqi Zheng1, Guixiang Ma2
1Key Laboratory of Trustworthy Distributed Computing and Service (MoE),
Beijing University of Posts and Telecommunications, China
2 University of Illinois at Chicago
{2019110945, zhangx, zjq1}@bupt.edu.cn, guixiang.ma@intel.com
当前存在的挑战
1.Though a set of rumor detection models have exploited the message propagation structural or temporal information, they seldom model them altogether to enjoy the best of both worlds
2.the dynamics of knowledge information associated with the comments are not involved
FND方法演变
解决方法
Dual-Dynamic Graph Convolutional Networks:DDGCN
一个GCN:evolving propagation graphs to capture both spatial and temporal structure of the messages as dynamic propagation representations
另一个GCN:evolving knowledge graphs associated with the messages to
learn dynamic knowledge representations
sequential fusing method:it consists of a sequence of temporal fusing units, and each unit would combine the two intermediate representations at the end of each time stage,and then passes the fused information to the successor unit
in a sequential manner
Problem Definition
Dynamic Graph Construction Module
Dynamic Propagation Graph Construction
Dynamic Knowledge Graph Construction
Dual-Dynamic GCN Module
A single Dual-Static GCN unit
A Temporal Fusion Unit
connecting the above two units
Rumor Classification Module
1.最后一层TF输出的是post node 和 comment node 的representations
使用mean-pooling operators 来整合节点表征信息
2.整合传播和知识表征 与 源推文节点表征
3.得到预测结果
4.损失函数 cross-entropy
实验
实验数据集
比较方法及结果
模型内部功能比较
传播图对结果的影响较大 = 传播图很重要