In adversarial training, there is one generator and one discriminator. The generator aims to generate fake data and “fool” the discriminator into thinking it’s real data, and the discriminator aims to tell fake data from real data. Here, data is assumed to have some prior distribution, and the GAE is used as the generator, so the GAE is forced to match the prior.

Improving GAE using Adversarial Training

Similar to GCN, some GAE methods incorporate nodes attributes into encoder and edge information in the training set is only used when training, so that the mapping function can be applied to nodes or graphs that are not seen in the training set.

Improving GAE using Inductive Learning

There are lots of similarity measures, such as L2-reconstruction loss, Laplacian eigenmaps, and KL divergence. There are different motivations for using them, but the study of their underlying difference still lacking.

Improving GAE using Similarity Measures

There are several perspective to improve a GAE model:

1)	Adversarial Training

2)	Inductive Learning

3)	Similarity Measures


University of Notre Dame

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)

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

This type of GAE is the most basic type of GAE. It views the adjacency matrix as features of nodes, and uses AE to reduce its dimension and obtain new representation. For example, Structure deep network embedding (SDNE) aims to let two nodes share similar representations if they have similar neighborhoods, and both first and second order nodes are preserved.

Original Graph Autoencoders

Variational Autoencoders is a type of GAE method that combine dimension reduction methods and generative models. For example, Deep Variational Network Embedding (DVNE) represents each node as a Gaussian Distribution and does dimension reduction by using the Wasserstein distance to promote nodes similarities.

Learn Before

Related

Learn After