Under node level graph RNN, each node is represented as a lower dimension vector, and the vector for each node is an aggregation of vectors of its neighbor node. And this aggregation is repeated multiple times to encode graph structure information in node representation. 

Node Level Graph RNN

Graph level RNN use RNN to capture graph level pattern, such as temporal pattern of dynamic graph and sequential pattern under different graph granularities. 

An example under this category is Dynamic Graph Neural Network (DGNN). DGNN can capture the temporal pattern of a dynamic graph. When a new edge is added to the graph, it uses LSTM to update the two newly connected nodes and their immediate neighbors.

Graph Level RNN

RNN is a class of neural network models that model sequential data. Examples include Gated Recurrent Units (GRU) or Long Short-Term Memory (LSTM) model. RNN can be adapted to graph modeling with some moderate modification.

University of Notre Dame

According to model structure and training strategies, deep learning algorithms on graphs can be divided into 5 categories:

1)	Graph Recurrent Neural Networks

2)	Graph Convolutional Networks

3)	Graph Autoencoders

4)	Graph Reinforcement Learning

5)	Graph Adversarial Methods


Types of Deep Learning Algorithms on Graphs

Zhang, Z., Cui, P., & Zhu, W. (2020). Deep learning on graphs: A survey. IEEE Transactions on Knowledge and Data Engineering.

Deep Learning on Graphs: A Survey

Graph Recurrent Neural Networks (RNN)

GCN uses convolution operation and readout operation to grasp node level and graph level patterns.

Graph Convolutional Networks (GCN)

Autoencoder (AE) is a type of neural network that aims to learn encodings of unlabeled data, and it’s evaluated by mapping the generated encoding back to the original input (decoder). It can be used to learn graph representation with moderate modification. Below are several types of it:

1)	Original Graph Autoencoders

2)	Graph Variational Autoencoders

3)	Other Improvements.


Graph Autoencoders (GAE)

Reinforcement learning is a deep learning method that learns from feedback and is effective in learning interactive games such as hide and seek. It can also be applied in graph setting. For example, Graph Convolutional Policy Network (GCPN) models graphs with a generative model, and regards this model as RL agent, treats link prediction as agent action, using domain-specific rewards, and uses GCN to learn node representation.                                     

Graph Reinforcement Learning (RL)

Adversarial Methods are a promising field and have drawn lots of attention. And there are mainly two ways of applying adversarial methods to graphs:

1) Adversarial Training

2) Adversarial Attacks


Learn Before

Related

Learn After