How to split a heterogeneous graph for link_pred targeting specific edge_type?

[SohaK]

Hi
I want to split a heterogeneous graph into train, dev, and test set for link_pred, but I need to have the positive and negative instances of the edges/links from a specific edge_type. Predicting that kind of edge is the problem description, other edges are just informative. Is there a way to do that with split?

Thanks,
Soha

[anniekmyatt]

I might be wrong but I think deepsnap’s splitting functionality will give you positive and negative supervision edges for all edge types in a hetero graph. If you want to specify one edge type for your supervision edges maybe have a look at pytorch-geometric’s RandomLinkSplit, see for example how it’s used here. In the same folder there is also a hetero_link_pred example which I contributed, so I hope it doesn’t seem like showing off that I share it😅, it just seems fairly relevant to what you’re trying to do.

[SohaK]

Thanks @anniekmyatt both RandomLinkSplit functionality and the code example are very useful for me.