超图
计算机科学
人工神经网络
邻接矩阵
图形
邻接表
节点(物理)
理论计算机科学
平滑的
人工智能
数据挖掘
数学
算法
组合数学
结构工程
工程类
计算机视觉
作者
Kunhao Li,Zhenhua Huang,Zhaohong Jia
标识
DOI:10.1109/tnse.2023.3243058
摘要
Graph neural networks have been widely studied and applied to node classification in graph-format datasets in recent years. Traditional graph neural networks mainly consider the adjacency characteristics of nodes and fail to learn rich role representations of nodes. Existing role representations methods of nodes are mostly in unsupervised approaches, resulting in unsatisfactory performance in downstream tasks. A graph can be reorganized as a hypergraph, in which the role characteristics of nodes are more intuitively represented. Based on this, we propose a role-aware hypergraph neural network (RAHG) that utilizes hypergraphs and an attention mechanism to fuse nodes' role and adjacency representations. A residual network is also applied to relieve the smoothing problem between layers in the model. The model adjusts the weights on the role and adjacency representations according to the characteristics of the graphs. RAHG significantly improves the prediction performance compared with existing graph neural networks on seven datasets, with accuracy increased by up to 12.1% on the node classification task in graphs.
科研通智能强力驱动
Strongly Powered by AbleSci AI