This reporsitory is an unofficial implementation of [2] GVAE.

An autoencoder is a neural network that can be used to encode and decode data (figure 1). in such a network, the encoder is a neural network that takes an input and outputs a hidden representation (latent), and the decoder is a neural network that takes the hidden representation and tries to reconstruct the original input.

A graph autoencoder is a graph neural network that can be used to encode graph structured data (figure 2). The encoder can be expresed as a graph neural network [1] that takes an input graph and outputs a hidden representation (latent). The decoder is a simple dot product that takes the hidden representation and tries to reconstruct the adjacency matrix.

To run experiments, you can use the train.py python file.

python train.py --dataset cora --hidden-size 32 --code-size 16 --lr 0.01

available datasets are: cora, pubmed, citeseer

[1] Kipf, Welling, et al. (2016). Semi-Supervised Learning with Graph Convolutional Networks.
[2] Kipf, T., & Welling, M.. (2016). Variational Graph Auto-Encoders.

• Implement the graph autoencoder
• Run experiments on citation networks
• Implement the graph variational autoencoder
• Write the GVAE section
• Add diagrams to the README (VGAE)
• Implement Dense GAE (not pushed to github yet)
• Add KL divergence to the loss function