Having RandomLinkSplit to generate negative samples that satisfy a condition

[yorgancioglukaan]

Hi! I am following the link prediction on heterogenous network example from here with my own data. However, I would like to use negative edges that are adjacent to a subset of nodes. Specifically, I want to ensure that train/test/validation data only considers ["user", "rates", "movie"] negative edges from specific set of "user" nodes of my own choosing. Positive edges are not a problem as they intrinsically satisfy this condition.

I think for this, I need to make changes to RandomLinkSplit where it samples negative edges. Is this a good starting point? Or is there a better way of doing this that I cant think of?

Thanks in advance!

Correction: Looks like LinkNeighborLoader is doing the negative sampling in this example.

[yorgancioglukaan]

If anyone else needs to do something like this, I ended up preparing train/test/val_data objects myself bypassing RandomLinkSplit and feeding those into LinkNeighborLoader.