图神经网络架构简要笔记

时间：2023.10.11

合集

• high way network，https://arxiv.org/pdf/1505.00387.pdf
• high way GCN, https://arxiv.org/pdf/2004.04635.pdf
• ResGCN, https://browse.arxiv.org/pdf/1904.03751.pdf
• × pmlr, On Over-Squashing in Message Passing Neural Networks: The Impact of Width, Depth, and Topology, https://proceedings.mlr.press/v202/di-giovanni23a.html
• icml'18, jump knowledge net, https://arxiv.org/abs/1806.03536
• mix-hop， https://arxiv.org/pdf/1905.00067.pdf
• × High-order graph attention network, https://www.sciencedirect.com/science/article/pii/S0020025523002426， https://dl.acm.org/doi/abs/10.1016/j.ins.2023.02.054
• × AAAI‘23，higher-order attention networks， https://arxiv.org/abs/2206.00606v1
• × graph ordering network， https://arxiv.org/pdf/2204.05351.pdf
• DMKD, Hierarchical GCN, https://link.springer.com/article/10.1007/s10618-022-00890-9
• NIPS'20, IGNN, https://browse.arxiv.org/pdf/2009.06211.pdf

highway network

graph highway network

与highway network同

GCN

略过不讲

ResGCN

CVPR2019，ResGCN, https://browse.arxiv.org/pdf/1904.03751.pdf

架构

1. 类似ResNet做残差
   
2. 提出graph上的空洞卷积dilated conv。
   

结果

目的

计算机视觉-segmentation

GNN理论

PMLR'23, On Over-Squashing in Message Passing Neural Networks: The Impact of Width, Depth, and Topology, https://proceedings.mlr.press/v202/di-giovanni23a.html

概念

width是中心点与周围邻居的数量；depth是graph上2个node相聚的距离（跳数）；topology，就是图的拓扑结构。文章以节点分类任务深入研究了此3个因素对过平滑的影响。

结论

Jump Knowledge Network

JKNet，ICML2018，老文章了，提出了个这个。

Mix-hop

ICML2019，也是文章，提出了这个。

hierarchical GNN

目的：node representation classification、link prediction

IGNN

NIPS'20
目的：解决recurrent gnn的不足
实验：node classification