use the whole graph adjacency matrix for link prediction task?

[LeeJunHyun]

https://github.com/snap-stanford/ogb/blob/master/examples/linkproppred/collab/gnn.py#L106

it seems that gnn model takes the whole adjacency matrix (data.adj_t).

but as far as I know, in the standard setting, gnn takes an incomplete set of edges (split_edge['train']['edge']) and predicts the rest (split_edge['valid'] and split_edge['test']).

Should I fix it? or could you please give me some reference for this setting?

I really appreciate the great commitment of you all.

[weihua916]

Hi, the graph object only contains training edges, so there is no information leak.

[LeeJunHyun]

Hi @weihua916 ,

Thanks for your reply.

Do you mean that the graph object (PygLinkPropPredDataset()) only has training edges?

PygLinkPropPredDataset() and data.adj_t have same number of edges.

Then how did you extract test and valid edges?

In dataset code, https://github.com/snap-stanford/ogb/blob/master/ogb/linkproppred/dataset_pyg.py#L67
they are just loaded from the split path.

dataset
PygLinkPropPredDataset()

dataset[0]
Data(edge_index=[2, 2358104], edge_weight=[2358104, 1], edge_year=[2358104, 1], x=[235868, 128])

data.adj_t
SparseTensor(row=tensor([ 0, 0, 0, ..., 235867, 235867, 235867]),
col=tensor([ 20649, 21913, 46512, ..., 230251, 230251, 235583]),
val=tensor([1., 1., 1., ..., 1., 3., 2.]),
size=(235868, 235868), nnz=2358104, density=0.00%)

split_edge['train']['edge'].shape
torch.Size([1179052, 2])

[LeeJunHyun]

This is what I know as a standard-setting, so could you please give me some other references that you have?
(the below image is from GRL book)

[weihua916]

1179052 * 2 = 2358104.

[LeeJunHyun]

Maybe there is something that I missed,

so is adj_t equal to split_edge['train']['edge']?

[weihua916]

https://ogb.stanford.edu/docs/nodeprop/
Note: For undirected graphs, the loaded graphs will have a doubled number of edges because we add the bidirectional edges automatically.

[LeeJunHyun]

I thought that valid and test edges are from adj_t.

then could you tell me where valid and test edges are from?

[weihua916]

They are from split_edge['valid']['edge'] and split_edge['test']['edge']

[LeeJunHyun]

Oh, I mean, where split_edge['valid']['edge'] and split_edge['test']['edge'] are from?

In dataset code, https://github.com/snap-stanford/ogb/blob/master/ogb/linkproppred/dataset_pyg.py#L67
they are just loaded from the split path.
(there are train.pt, valid.pt, and test.pt in the split path)

how did you create train.pt, valid.pt, and test.pt?

[LeeJunHyun]

And I performed the validation to find overlapped edges between data.adj_t and split_edge['test']['edge'].
There are overlapped edges in both train and test sets.
please let me know if there is another point that I'm still missing.

        row, col, _ = data.adj_t.coo()
        total_edge = torch.stack([row,col],dim=0).t()

        for test_edge in split_edge['test']['edge']:
            overlap_validation = (test_edge==total_edge).sum(dim=1)==2
            if overlap_validation.max() >0:
                overlap_idx = overlap_validation.nonzero().squeeze()
                print(f'Test edge: {test_edge}')
                print(f'Train edge: {total_edge[overlap_idx,:]}')
                print('*'*20)

[LeeJunHyun]

The results:

Test edge: tensor([105700, 201535])
Train edge: tensor([[105700, 201535],
[105700, 201535],
[105700, 201535],
[105700, 201535],
[105700, 201535]])

---

Test edge: tensor([159698, 220004])
Train edge: tensor([159698, 220004])

---

Test edge: tensor([ 10737, 220004])
Train edge: tensor([[ 10737, 220004],
[ 10737, 220004],
[ 10737, 220004]])

---

Test edge: tensor([106042, 220004])
Train edge: tensor([[106042, 220004],
[106042, 220004]])

---

Test edge: tensor([131588, 161470])
Train edge: tensor([[131588, 161470],
[131588, 161470]])

---

Test edge: tensor([117925, 112050])
Train edge: tensor([[117925, 112050],
[117925, 112050],
[117925, 112050],
[117925, 112050]])
...

[LeeJunHyun]

When I set total_edge=split_edge['train']['edge'], there are also overlapped edges (between split_edge['train']['edge'] and split_edge['test']['edge']).

[weihua916]

The overlapping edges are expected in ogbl-collab. See https://ogb.stanford.edu/docs/linkprop/#ogbl-collab.

[LeeJunHyun]

Because it is split by time (year). It makes sense now.

Thanks for your help!