Help me understand LinkNeighborLoader(disjoint=True)

[brovatten]

Hi,

I'm using the LinkNeighborLoader for link prediction for some edges. To avoid data leakage, I want the two subgraphs generated for a link to be disjoint, meaning there is no path between them. But when I set disjoint=True, the returned Data object contains an edge_label_index size (2,2). Shouldn't the edge_label_index just contain my input link?

As I use batch_size=1 and the output of .batch is all zeros for every result, I presume the disjoint option is for multiple batches, and that LinkNeighborLoader guarantees no data leakage, but I wanted to ask here to make sure.

These are my arguments:
LinkNeighborLoader(data=data, edge_label_index=eli, edge_label=el, num_neighbors=[80,40], batch_size=1, is_sorted=True, shuffle=False, disjoint=True)

Thanks in advance!

[rusty1s]

This is currently only possible by ensuring that data.edge_index and data.edge_label_index do not contain any shared values. You can achieve this via RandomLinkSplit for example via the disjoint_train_ratio argument.

[brovatten]

Actually no, because the indices you see in edge_label_index is are reindexed. The indices in there are mapping to the indices in n_id containing the original node ids. You can see here how to retrieve them:

original_nodes = [2, 5, 7, 9] 
reset_edge_index = torch.tensor([[0, 1, 2],[1, 2, 3]])  
mapped_edge_index = torch.tensor(original_nodes)[reset_edge_index]

tensor([[2, 5, 7],
              [5, 7, 9]])

[wgeul]

@brovatten thanks, that's helpful. I am working with a heterogeneous graph but I think the same principle will apply.

[brovatten]

Okay I think I figured it out. Please correct me if I am wrong @rusty1s

edge_label_index will produce double the amount of edges input because it will disjoint both nodes in a link. This means that there now is link in the edge_label_index that connects it with the subgraph of the head, and a second link that connects it with the subgraph of the tail.

The disjoint argument will only guarantee no leakage on a link to link basis. Meaning if we batch several links, there is no guarantee that they information will not leak through subgraphs created by other links in the batch. This needs to be accounted for manually.

Edit: After examining more I am more confused. How does this input become this output:

tensor([
[ 756111, 1083433, 2155848, 4731623, 5086442],
[1244459,  181071, 2119679,  811872,  323703]])
        
tensor([
[ 756111, 1083433, 2155848, 4731623, 5086442, 1244459,  181071, 2119679, 811872,  323703],
[ 973070, 3498205, 4163093, 3499019, 5348527, 4564240, 3563536, 4720508, 5252297, 4386027]])

I do notice the first 5 heads in the outputs are heads in the input and the last 5 heads in the outputs are tails in the input. But what does the tails in the output constitute?

[rusty1s]

Can you show me an example of how you get this output? You may get an increased size back for edge_label_index due to negative sampling, but positive tails shouldn't move to heads instead.

[rusty1s]

Please also see #7943 where we test that the local and global edge_label_index match.