search

pyg-team / pytorch_geometric

Graph Neural Network Library for PyTorch
https://pyg.org
MIT License
20.52k stars 3.57k forks source link

Make train/val/test splits disjoint sets in RandomLinkSplit #9397

Closed mk-61 closed 3 weeks ago

mk-61 commented 3 weeks ago

It seems train split is erroneously used instead of validation split, and also I don't understand what train_val_edges variable is used for. So this PR is an attempt to fix this. But some other tests are failing, so consider this as WIP, for now...

rusty1s commented 3 weeks ago

This is actually intended. We want to use the edges for training and validation in the test phase for message passing. Note that edge_label_index (the supervision) edges are disjoint, but it is not a requirement for edge_index.

mk-61 commented 3 weeks ago

Right, makes sense, thanks for clarification!