Data leakage issue when using NeighborSampler?

[Dennis-Tsai]

While using NeighborSampler, we set our selected train_id and valid_id (without overlapping) as the node_idx for creating the mini-batch training dataloaders as followed:

train_loader = NeighborSampler(
    edge_index,
    node_idx=train_idx,
    sizes=[-1, -1, -1],
    batch_size=64,
    shuffle=True,
    num_workers=12,
    drop_last=True,
)

val_loader = NeighborSampler(
    edge_index,
    node_idx=val_idx,
    sizes=[-1, -1, -1],
    batch_size=64,
    shuffle=True,
    num_workers=12,
    drop_last=True,
)

However, once we looped through the train_loader as shown below, the n_id, the nodes involved in the training computation, have some overlapping with our val_idx:

for batch_size, n_id, adjs in train_loader:
    .....

The valid/test node data can be accessed in training process. I wonder if it is kinda data leakage problem?
Please let me know if I misunderstood any part. Thanks!

[wsad1]

This is the expected behaviour.
The n_id is the id of all nodes involved in message passing this can overlap across train/val set since they can have overlapping 3 hop neighbourhood. But only the train_idx nodes will be used for supervision.

[Dennis-Tsai]

Thanks for the prompt reply. I'm just thinking if there is any concern that the model could learned some knowledge from the valid dataset's node features in training process. In this case we could expect to gain a better performance when evaluating by this valid dataset.

Maybe adding a mask mechanism to prevent those overlapping would be more reasonable?

[rusty1s]

If you want to apply your model in an inductive rather than transductive learning scenario, you can simply filter it before passing it to the loader via train_data = data.subgraph(data.train_mask).