Hi everyone, I am new to torch_geometric and the GNN world.
If my network doesn’t have a feature matrix X but I want my model to rely only on the topology of the network what are the best ways to initialise this feature matrix X.
I have tried: torch.eye(N,N) and torch.rand(N,N). I noticed that the model performs better when I have the identity matrix as the feature matrix rather than giving random entries.
However my network is pretty big (about 30,000 nodes) and having that amount of features convoluted into a say 128 latent space makes it impossible to converge (the model is not learning anything).
Do you have any ideas on what would be the best solution here, or any papers that I could read regarding this problem?
Thank you!