Bug in RandomLinkSplit

[dpaysan]

Dear all,

I had some issues with some dimension mismatches and thereby saw that after determining the random split indices in the RandomLinkSplit finally the data is split not as I had expected. In the RandomLinkSplit class the data splits actually do not use the sampled validation and test edges but reuse the train edges for the validation data and a combination of the training and validation edges for the test data. I have copied the respective part of the code below which can be found here :

    train_edges = perm[:num_train]
    val_edges = perm[num_train:num_train + num_val]
    test_edges = perm[num_train + num_val:]
    train_val_edges = perm[:num_train + num_val]

    # Create data splits:
    train_data = self._split(data, train_edges)
    val_data = self._split(data, train_edges)
    test_data = self._split(data, train_val_edges)``

Could somebody explain why that is?

Thanks a lot in advance!

[rusty1s]

data.edge_index refers to the edges that are used for message passing. As such, during training and validation, you are allowed to propagate information based on the training edges, while during testing, you can propagate information based on the union of training and validation edges.

For evaluation, the data.edge_label_index holds a batch of positive and negative samples that should be used to evaluate your model on.

So:

• data.edge_index should be solely used for message passing
• data.edge_label_index should be used for evaluation

[dpaysan]

That makes sense. Thanks!