Query on Variational Graph Auto-Encoder

[shrimonmuke0202]

Hi,
My question is, how can I split the latent representation of GVAE into two parts? One is for nodes, and the other is for edges. Is this splitting of the latent representation of GVAE possible? I am waiting for your positive reply.

[rusty1s]

Can you clarify what you want to try to do? Note that VGAE only computes latent representations for nodes, and uses them to reconstruct the adjacency matrix. It should be nonetheless possible to extend upon this scheme to compute both node and edge representations using a GNN, and using both to reconstruct the input.

[shrimonmuke0202]

I want to reconstruct node and edge features from the latent representation. Thus, I have to use a linear layer on the latent representation to get node and edge embedding. Also, can you give me an example of how I can apply GVAE to the ZINC dataset i.e., in the molecular graph generation task?

[rusty1s]

I sadly do not have an implementation of this, but happy to help out if you have any questions on how to do so.

[shrimonmuke0202]

In this example, VGAE is applied on a single graph. My query is how to process the same thing on batched graphs (multiple graphs in a single batch)?

[rusty1s]

How about

dataset = MyDataset(..., transform=RandomLinkSplit(...))
train_data_list = []
for train_data, val_data, test_data in dataset:
    train_data_list.append(train_data)

train_loader = DataLoader(train_data_list, batch_size=..., shuffle=True)

and then apply the pipelines as given (while iterating over all data inside train()). Importantly, you may need to provide your own negative samples (via batched_negative_sampling) here.

[shrimonmuke0202]

Thanks for your comments; this will be very helpful for me. If there is any problem during the implementation of the code, then I will ping in the following thread.