Node Embedding using GAE

[elbasir]

Hi,

My understanding of GAE is the capability to generate an embedding which represents an entire graph, but assume I have a graph G with N nodes without any labels, I want to extract an embedding for each node in G. Can you guide me on how to get a node embedding using GAE or if there is another unsupervised method which can handle this?

Thanks,

[rusty1s]

GAE will produce node embeddings such that nodes that should be connected to each other have a similar embedding while nodes that are far away from each other should also be far away in the embedding space (similar to Node2Vec but based on GNN as encoder). As such, GAE already returns node-level embeddings as part of its encoder.

[elbasir]

Thanks for the answer. Assume I have a set of proteins, if I generate a graph of each protein, I pass each graph to GAE and the embedding would be representing the entire protein graph, what about if I have PPI network and I want to generate embedding of each node "protein" here? I checked GAE examples, and what I noticed, it takes the entire graph and generate only one embedding, can you provide me an example how GAE will produce node embeddings in a such scenario as PPI network? I appreciate your assistance.

[rusty1s]

I think your understanding is slightly incorrect. As mentioned earlier, GAE will not embed a graph, but will put each node in an embeddding space. You can see this here where the encoder returns a matrix of shape [num_nodes, num_embedding_dim]. As such, GAE already does what you want.

[elbasir]

Excuse my incorrect understanding but now it is clear. So, if I want to have an embedding for the entire graph should I aggregate all the num_embedding_dim for all nodes in the graph?

[rusty1s]

What you can also try is Deep Graph Infomax (see examples/infomax.py), and compare it to the GAE approach.

[elbasir]

Does Deep Graph Infomax retrieve an embedding of an entire graph in this case or it follows the same learning of GAE? I will investigate both of them to have an unsupervised approach for generating a graph representation.

[rusty1s]

It actually can both - graph-level representations can be obtained through the summary vector.

[elbasir]

BTW, I checked examples folder but infomax.py does not exist but there are examples/infomax_inductive.py and examples/infomax_transductive.py, which one should be used?

[rusty1s]

Ah right. I suggest to look into the transductive one first.