图神经网络（GNN）1

图论基本概念

图（节点、边）、有向图与无向图、walk、trail、path、cycle、完全图、正则图、简单图、multigraph、pseudograph、hypergraph、degree等。

Outline

• Brief history and some graph application
• Some definitions about graph
• Taxonomy of GNN
• Tasks
• Graph convolution networks
• Graph attention networks
• Graph autoencoder
• Graph spatial-temporal networks
• Graph network (GN) block

Brief history

1998 LeNet - LeCun
2003 Language model - Benglo
2005 GNN - Franco Scarselli, Marco Gori, Gabriele Monfardini
2006 RBM - Hinton
2009 GNN - Marco Gori, Gabriele Monfardini, Franco Scarselli
2012 AlexNet - Hinton
2015 Deep learning - Hinton

Graph applications

• Social network
• Biological network (gene regulatory network, protein-protein interaction network)
• Computer graphics
• Traffic network
• Knowledge graph
  
  
  Social network
  
  
  PPI Network
  
  
  Knowledge graph
  
  PageRank
  

Definitions

• graph
• Adjacency matrix (edge exist ?)
  
  
• Degree matrix
  
• Laplacian matrix
  
• Symmetric normalized Laplacian matrix
  
• Graph signals and labels
  
  
• GNN and network embedding
  

Taxonomy of GNN

Tasks

node-level tasks

• node regression, node classification
• graph convolution layer gives node’s latent representations

edge-level tesks

• link prediction, edge classification
• additional function would take two nodes’ latent representations as input of graph convolution layer

graph-level (global-level) tasks

• graph classification
• graph pooling gives the reduced graph information

Semi-supervised learning for node classification

Supervised learning for graph classification

Unsupervised learning for graph embedding
Convolution on graph