Train/Test leak in RandomNodeSplit?

[AdarshMJ]

Hi @rusty1s I am trying to perform node classification and I use the following transform on the cora dataset -

transform2 = RandomNodeSplit(split="test_rest",num_splits = 10)
data  = transform2(data)
print(data)

If I do -

train_nodes = data.train_mask.nonzero(as_tuple=True)[0].cpu().numpy()
test_nodes = data.test_mask.nonzero(as_tuple=True)[0].cpu().numpy()
leakage_nodes = np.intersect1d(train_nodes, test_nodes)
if len(leakage_nodes) > 0:
        print(f"Warning: Found {len(leakage_nodes)} nodes in both the training and test sets.")
    else:
        print("No leakage detected.")

I get -

Found 1045 nodes in both the training and test sets

Am I using the transform incorrectly?

[rusty1s]

Don't you need to check every the splits in isolation?

import numpy as np

from torch_geometric.datasets import Planetoid
from torch_geometric.transforms import RandomNodeSplit

dataset = Planetoid('/tmp/Cora', name='Cora')
data = dataset[0]
transform2 = RandomNodeSplit(split="test_rest", num_splits=10)
data = transform2(data)

for i in range(10):
    train_nodes = data.train_mask[:, i].nonzero(as_tuple=True)[0].cpu().numpy()
    test_nodes = data.test_mask[:, i].nonzero(as_tuple=True)[0].cpu().numpy()

    leakage_nodes = np.intersect1d(train_nodes, test_nodes)
    if len(leakage_nodes) > 0:
        print(
            f"Warning: Found {len(leakage_nodes)} nodes in both the training and test sets."
        )
    else:
        print("No leakage detected.")

No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.
No leakage detected.

[AdarshMJ]

Yes that makes sense. Thank you :)