异质图神经网络在属性缺失数据上的应用
Application of Heterogeneous Graph Neural Networks on Attribute Missing Data
李振军 1赵华 1刘祖军 1陶周天 1杨斌 2谭卓 3黄嘉琦 3邢颖3
作者信息
- 1. 智慧足迹数据科技有限公司,北京 100033
- 2. 中国联通研究院,北京 100048
- 3. 北京邮电大学 人工智能学院,北京 100876
- 折叠
摘要
图数据挖掘是异质图神经网络(Heterogeneous Graph Neural Network,HetGNN)模型的主要任务,HetGNN用图嵌入技术对图结构数据进行表示和计算.然而,在大多数实际场景中,异质图提供的信息并不完整.对比了 3 种图数据挖掘模型——元路径聚合图神经网络(Metapath Aggregated Graph Neural Network,MAGNN)、HetGNN和异质图注意力网络(Heterogeneous Graph Attention Network,HAN)模型.在 2 个公开可用的异质图数据集上进行了实验,结果表明,在节点分类的数据挖掘任务中,MAGNN模型的性能优于HAN和HetGNN,证明了其有效性.
Abstract
Graph data mining is the main task of Heterogeneous Graph Neural Network(HetGNN)model,which uses graph embedding techniques to represent and compute the graph structure data.However,in most practical scenarios,the information provided by heterogeneous graphs is incomplete.Three kinds of graph data mining models including Metapath Aggregated Graph Neural Network(MAGNN),HetGNN and Heterogeneous Graph Attention Network(HAN)model are compared.Experiments are carried out on two publicly available heterogeneous graph datasets.The results show that the performance of MAGNN model is better than that of HAN and HetGNN in the data mining tasks of node classification,which proves its effectiveness.
关键词
异质图/神经网络/数据挖掘Key words
heterogeneous graph/neural network/data mining引用本文复制引用
出版年
2024
NETL
NSTL
万方数据