Using RandomLinkSplit + NeighborLoader.

[Yoloniverse]

To do a link prediction task for heterodata, I want to use both RandomLinkSplit and NeighborLoader. My dataset does not have masking. Therefore it is easier to use RandomLinkSplit. But when I want to use NeighborLoader, I should provide input_nodes in a form of input_nodes=('node_type', hetero_data['node_type'].train_mask). How can I provide train_mask here when RandomLinkSplit does not return train_mask? Maybe it would be better to provide the mask when doing RandomLinkSplit.

And for now as far as I understand, the NeighborLoader is based on node level tasks. Would you add update about NeighborLoader for link prediction? Thank you.

[rusty1s]

Please see and comment on my newly created issue in #4026, which tries to integrate link-level prediction support to NeighborLoader. I will try to integrate this feature ASAP. Currently, this can only be achieved by making use of some workarounds (also linked in the issue).

Also note that it is well possible to pass either masks or indices to input_nodes.

[Yoloniverse]

I really appreciate you to create the issue #4026! Thank you.