摘要
目前大多数图卷积(GCN)关注于提取局部特征信息,忽略了全局特征,使得标签不能有效地传播到整个图上。对此,设计一种可以用于半监督节点分类任务的深度池化对偶图神经网络(DPDNN)。该网络引入池化机制,将结构相似的节点聚合为超节点,扩大节点的接收域。通过随机游走得到图全局信息的潜在表示,使用随机游走模型与GCN进行协同训练,从而补充GCN模型在获取整个图拓扑结构信息上的能力。实验结果表明,该网络模型与现有方法相比提高了分类精度,在少量数据标记时效果更为明显。
At present,most graph convolution(GCN)focuses on extracting local feature information while ignoring global features,so that tags cannot be effectively propagated to the entire graph.This paper designed a deep pooled dual graph neural network(DPDNN)that can be used for semi-supervised node classification tasks.This network introduced a pooling mechanism to aggregate nodes with similar structures into super nodes and expand the receiving field of each node.Through random walk,DPDNN obtained the potential representation of the global information of the graph,and by using the random walk model to coordinate training with GCN,the ability of the GCN model to acquire the whole graph topology information was supplemented.The experimental results show that compared with the existing methods,this network model improves the classification accuracy,and the effect is more obvious in a small amount of data markers.
作者
薛磊
农丽萍
张文辉
林基明
王俊义
Xue Lei;Nong Liping;Zhang Wenhui;Lin Jiming;Wang Junyi(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,Guangxi,China;School of Telecommunication Engineering,Xidian University,Xi’an 710071,Shaanxi,China;College of Physics and Technology,Guangxi Normal University,Guilin 541004,Guangxi,China)
出处
《计算机应用与软件》
北大核心
2021年第10期153-158,163,共7页
Computer Applications and Software
基金
国家自然科学基金项目(61966007)
“认知无线电与信息处理”省部共建教育部重点实验室基金项目(CRKL180201,CRKL180106)。
关键词
图卷积网络
图信号
图池化
半监督学习
Graph convolutional network
Graph signal
Graph pooling
Semi-supervised learning
