How to understand the disjoint_train_ratio parameter in RandomLinkSplit?

[ki-ljl]

When I use RandomLinkSplit to split the dataset, I get results that I cannot understand:

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
transform = T.Compose([
    T.NormalizeFeatures(),
    T.ToDevice(device),
    T.RandomLinkSplit(num_val=0.1, num_test=0.1, is_undirected=True,
                      add_negative_train_samples=False),
])
dataset = Planetoid('data', name='CiteSeer', transform=transform)
train_data, val_data, test_data = dataset[0]

I get the following results:

Data(x=[3327, 3703], edge_index=[2, 7284], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[3642], edge_label_index=[2, 3642])
Data(x=[3327, 3703], edge_index=[2, 7284], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[910], edge_label_index=[2, 910])
Data(x=[3327, 3703], edge_index=[2, 8194], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[910], edge_label_index=[2, 910])

In my understanding, edge_index=[2, 7284] in train_data is training supervision edges, that is, edges used for GNN message passing and aggregation. edge_label_index=[2, 3642]) is the training message edges, that is, the edges used for supervised learning.

Now we try to set disjoint_train_ratio to 0.5 and we will get the following result:

Data(x=[3327, 3703], edge_index=[2, 3642], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[1821], edge_label_index=[2, 1821])
Data(x=[3327, 3703], edge_index=[2, 7284], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[910], edge_label_index=[2, 910])
Data(x=[3327, 3703], edge_index=[2, 8194], y=[3327], train_mask=[3327], val_mask=[3327], test_mask=[3327], edge_label=[910], edge_label_index=[2, 910])

It can be found that both edge_index and edge_label_index in train_data have become half of the previous ones. The explanation of the disjoint_train_ratio parameter in the PyG documentation is as follows:

If set to a value greater than 0.0, training edges will not be shared for message passing and supervision. Instead, disjoint_train_ratio edges are used as ground-truth labels for supervision during training. (default: 0.0)

Based on the above explanation I am confused by the previous output.

In my understanding if disjoint_train_ratio=0.0 means edge_label_index will be included in edge_index, if disjoint_train_ratio=0.5 means edge_label_index will not be included in edge_index (and it is), but I am curious why edge_index is also reduced half?

[rusty1s]

IMO, it makes sense that both are reduced by half. Previously, message passing edges and supervision edges were shared, while they are not disjoint in a 50:50 ratio. Does that make sense?

[ki-ljl]

Thank you for your prompt reply. Forgive me for another stupid question: is there any meaning to keep the ratio of edge_index and edge_label_index in train_data at 2:1? Is it to make the number of the two consistent after negative sampling? Or for other reasons?

[rusty1s]

The ratio is 2:1 in your case because you are operating on an undirected graph, that means that bidirectional edges are used for message passing, while only one direction of edges is used for supervision (since one typically uses a symmetric decoder).

[ki-ljl]

I see, thank you very much for your reply.

[sakunaharinda]

Hi, Can you please explain to me the terms training edges, supervision edges, and message passing edges and how they relate?