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
| Dataset | AUC | AP |
|---|---|---|
| Cora | 92.74 | 93.38 |
| Citeseer | 91.64 | 92.44 |
| Pubmed | --.-- | --.-- |
| Dataset | AUC | AP |
|---|---|---|
| Cora | 91.4 | 92.6 |
| Citeseer | 90.8 | 92.0 |
| Pubmed | 94.4 | 94.7 |
[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