Graph-level autoencoder about gae

[tkipf]

Hi, thanks for getting in touch. These are very good questions indeed.

1. In this work, we focused on auto-encoding the graph structure alone. We provide an option to condition the encoder on a set of node features. It is quite straightforward to extend this model to also decode node features or edge features (a simple MLP decoder will do in this case) - this could potentially improve predictive performance on a number of tasks.
2. If I understand correctly, you would like to learn latent representations of graphs instead of nodes and then cluster them. GAEs only support learning latent representations for nodes in a graph. Auto-encoding multiple graph instances comes with some difficulties. Since the set of nodes is unordered, reconstructing both nodes and edges from a fixed-size latent graph representation is challenging as it would involve evaluating all possible node orderings if done naively.

Hope this helps.

Thomas

[patrafter1999]

Thanks for your explanation, Kipf.