Training GNN on entire adj instead of training_set_adj

[nabihach]

I'm looking at this line of code: https://github.com/snap-stanford/ogb/blob/c8f0d2aca80a4f885bfd6ad5258ecf1c2d0ac2d9/examples/linkproppred/collab/gnn.py#L107

My question is, why are we training the GNN model on the entire adjacency matrix instead of just the training set's adjacency matrix? If the GNN can see the val and test edges, that information will be incorporated into the GNN's hidden states. Therefore when we send those hidden states to the LinkPredictor to predict presence/absence of edges, wouldn't that be considered data leakage?

[weihua916]

Great point. data.adj_t only contains training edge information.

[nabihach]

@weihua916 Can you please explain how that is so? I am seeing in Lines 221--225 that the edges are not being split when creating the variable data:

    dataset = PygLinkPropPredDataset(name='ogbl-collab')
    data = dataset[0]
    edge_index = data.edge_index
    data.edge_weight = data.edge_weight.view(-1).to(torch.float)
    data = T.ToSparseTensor()(data)

[weihua916]

data by default only contain training edges. The val/test edges can be accessed as

split_edge = dataset.get_edge_split()

[nabihach]

Got it. Thanks!