Example for link prediction validation/test time

[garyhsu29]

I want to use GraphSage to train a node embedding from a huge custom dataset (~17GB) and it's in one graph.
In my setting, there is no labeled data, so I could use link prediction to achieve my goal, but when I look into the example of GraphSage unsupervised setting https://github.com/pyg-team/pytorch_geometric/blob/master/examples/graph_sage_unsup.py
But in the example, the link prediction only applies to training time, but how could I do it during the validation/test time as well?
I checked RandomLinkSplit but it's still not very clear to me. Does anyone have related experience?

[rusty1s]

What do you wanna do during evaluation? If your goal is to simply obtain node embeddings of test/val nodes, you can pass these nodes to a separate NeighborLoader and run your model forward procedure on these mini-batches.

[garyhsu29]

Thank you for replying!
That says I want to also perform a linked prediction task on validation/test time.
Should I split the data before I passed to LinkNeighborLoader

train_loader = LinkNeighborLoader(
    data,
    batch_size=256,
    shuffle=True,
    neg_sampling_ratio=1.0,
    num_neighbors=[10, 10],
)

or should I do it after the LinkNeighborLoader

[rusty1s]

You would need to create a new loader that iterates over all validation edge_label_index, e.g.:

val_loader = LinkNeighborLoader(
    data=val_data,
    num_neighbors=[20, 10],
    edge_label_index=val_edge_label_index,
    edge_label=val_edge_label,
    batch_size=512,
    shuffle=False,
)