ICML'21 | 五篇图神经网络论文精选

国际机器学习大会（ International Conference on Machine Learning, ICML）是由国际机器学习学会（IMLS）主办的年度机器学习国际顶级会议，位列CCF-A，2021年论文接收率为21.48％。今天为大家带来ICML 2021五篇关于图神经网络过平滑问题与深度图神经网络与的文章分享。

1研究背景

深度图神经网络的训练过程非常艰难，需要大量的计算开销和存储空间。此外，深度图神经网络不仅容易带来梯度消失和过度拟合问题，还存在过平滑问题：图神经网络的训练过程中，随着网络层数的增加和迭代次数的增加，每个节点的隐层表征会趋于收敛到同一向量并逐渐变得难以区分，从而导致性能急剧下降。(注：过平滑问题不仅存在于深度图神经网络，普通的浅层图神经网络也会存在这个问题)

2Oversmoothing (过平滑问题)

1GRAND: Graph Neural Diffusion

作者：Benjamin P. Chamberlain, James Rowbottom, Maria Gorinova, Stefan Webb, Emanuele Rossi and Michael M. Bronstein
单位：Twitter 等
论文链接：http://proceedings.mlr.press/v139/chamberlain21a.html

Highlight: We present Graph Neural Diffusion (GRAND) that approaches deep learning on graphs as a continuous diffusion process and treats Graph Neural Networks (GNNs) as discretisations of an underlying PDE.

2Directional Graph Networks

作者：Dominique Beaini, Saro Passaro, Vincent Letourneau, William L. Hamilton, Gabriele Corso and Pietro Lio
单位：InVivo AI & 英国剑桥大学等
论文链接：http://proceedings.mlr.press/v139/beani21a.html
代码链接：https://github.com/Saro00/DGN

Highlight: To overcome the limitation of oversmoothing, we propose the first globally consistent anisotropic kernels for GNNs, allowing for graph convolutions that are defined according to topologicaly-derived directional flows.

3Improving Breadth-Wise Backpropagation in Graph Neural Networks Helps Learning Long-Range Dependencies

作者：Denis Lukovnikov and Asja Fischer
单位：Ruhr University Bochum
论文链接：http://proceedings.mlr.press/v139/lukovnikov21a.html

Highlight: In this work, we focus on the ability of graph neural networks (GNNs) to learn long-range patterns in graphs with edge features.

3Depth（深度图神经网络）

4Optimization of Graph Neural Networks: Implicit Acceleration by Skip Connections and More Depth

作者：Lukovnikov, Denis and Fischer and Asja
单位：MIT & 马里兰大学 & 哈佛
论文链接：http://proceedings.mlr.press/v139/xu21k.html

Highlight: We take the first step towards analyzing GNN training by studying the gradient dynamics of GNNs.

5Training Graph Neural Networks with 1000 Layers

作者：Li, Guohao and M{"u}ller, Matthias and Ghanem, Bernard and Koltun, Vladlen
单位：Intel Labs & King Abdullah University等
论文链接：http://proceedings.mlr.press/v139/li21o.html

Highlight: In this work, we study reversible connections, group convolutions, weight tying, and equilibrium models to advance the memory and parameter efficiency of GNNs.