Question about input parameter

[vymao]

Hi,

I was wondering what exactly the --train_node_embedding parameter is for? What else besides the DRNL label and input node feature is inputted into the DCGNN?

[muhanzhang]

If you append --train_node_embedding, an additional unique embedding per node id is concatenated with DRNL label and node feature before fed to a GNN, and is trained together with GNN parameters. By default it is disabled, because it is not observed to improve SEAL's performance while making it not inductive.

[vymao]

I see, thanks. And do you retrain on the validation edge set to test on the final test edge set?

[muhanzhang]

For ogbl-collab, after we determined all the hyperparameters, we included validation edges in the input graph and retrain the model. This is as suggested by https://ogb.stanford.edu/docs/leader_rules/. For other datasets, we don't use validation edges in training.