How can I define the range of negative sampling in the LinkNeighborLoader?

[songsong0425]

Hi. Although I think that there would be a duplicated question, I have a question about the negative sampling in the LinkNeighborLoader.
Is there a way to select the negative samples from the pre-defined range of nodes?
For example, when I set the loader as below, it contains the unnecessary negative edge_label_index.

train_loader = LinkNeighborLoader(kf_train_data, edge_label_index=kf_train_data.edge_label_index, edge_label=kf_train_data.edge_label,
                                      batch_size=32, shuffle=True, neg_sampling_ratio=1.0, num_neighbors=[2,2], disjoint=True, neg_sampling="binary")

I guess that the unknown edges in the edge_index of the training dataset were selected randomly as the negative samples in the batch.
(But it's a little weird because the documentation mentioned the edge_label_index and edge_lablel in the neg_sampling )

neg_sampling (NegativeSampling, optional) – The negative sampling configuration. For negative sampling mode "binary", samples can be accessed via the attributes edge_label_index and edge_label in the respective edge type of the returned mini-batch. In case edge_label does not exist, it will be automatically created and represents a binary classification task

When I have a node list and want to sample the random negative edges avoiding the duplicated edges with true positive edge_label_index, how can I extract random negative samples?

print(node_list)
print(len(node_list))
# array([     0,      1,      3, ..., 101089, 101090, 101091])
# 19392

Currently, I created random edges from the node list before entering the loader, but since I have modified the dataset it is difficult to sample new negative samples during the epoch.

random_neg_edges = torch.stack([torch.tensor(random.sample(list(node_list), len(GraphList.edge_label))), 
                                    torch.tensor(random.sample(list(node_list), len(GraphList.edge_label)))], 0)
GraphList.edge_index = torch.cat([GraphList.edge_index, random_neg_edges], dim=1)
GraphList.edge_label = torch.cat([GraphList.edge_label, torch.zeros(len(GraphList.edge_label)).int()])

[rusty1s]

Sorry, I don't understand the question yet. Why is edge_label_index useless? edge_label_index refers to the set of edges we want to train against, and they are extended by negative edges as well. Positive and negative edges can then be discriminated via edge_label (same number of elements as positive/negative edges).

If you want to sample on your own, you can also take a look at utils.negative_sampling.

[songsong0425]

I'm sorry for the poor question. I hope the below figure helps your understanding.
(If it is wrong, please let me know)

In short, I only want to sample the negative edges from the unknown links between blue nodes.
But when I set the neg_sampling_ratio=1.0, the loader automatically sampled edges from the edge_index containing black nodes. (I need to calculate the probability and evaluation metrics only between blue nodes)
Can I prevent this phenomenon?

[rusty1s]

We currently lack a good interface to customize negative sampling. What you can try is passing a custom NegativeSampling class to your loader, which modifies sample to only sample blue nodes.