LinkNeighborLoader changes the lable of edge_label_index

[SheidaMajouni]

Hi,

I’m need help with a bipartite graph I’m working on. based on the nature of my problem, my graph consists of two distinct predefined types of edges: “message passing” and “supervised.” The supervised edges in my problem are already assigned binary labels. I want to use supervised_label = 0 to denote my negative samples. I have carefully prepared my train_data, val_data, and test_data.

I plan to employ LinkNeighborLoader to generate batches for my dataset. I understand that due to the nature of neighborhood sampling I can see edges that were not in my train_data. However, I’ve encountered an unexpected issue where the labels of some edges are altered, and I’m unsure why this occurs.

I’m focusing on some nodes from my dataset. For my train_data, the edge_label_index is:

tensor([[2, 6, 1, 8, 4, 5, 0, 9],
[2, 6, 1, 8, 4, 5, 0, 9]])

The corresponding edge_labels are:

tensor([1, 1, 1, 0, 0, 0, 0, 0])

Here, the label for the edge (6, 6) is clearly 1.

However, after executing my code, the edge_label_index for sample_data becomes:

tensor([[6, 4, 5, 7, 0, 1, 2, 3],
[6, 4, 5, 7, 0, 1, 2, 3]])

With the corresponding edge_labels of:

tensor([0, 0, 1, 0, 0, 1, 1, 0])

In this case, two new edges (7,7) and (3,3) have been added due to sampling, which is acceptable. However, the label for edge (6,6) has been changed to 0. and why it didn't include edges of (9,9) and (8,8) which are in train edge_label_index?

Could someone clarify if my interpretation is correct? Why would the label change?

from torch_geometric.loader import LinkNeighborLoader

# Define seed edges:
edge_label_index = train_data["recipient", "match", "donor"].edge_label_index
print(edge_label_index) # ==> tensor([[2, 6, 1, 8, 4, 5, 0, 9], [2, 6, 1, 8, 4, 5, 0, 9]])

edge_label = train_data["recipient", "match", "donor"].edge_label
print(edge_label) # ==> tensor([1, 1, 1, 0, 0, 0, 0, 0])

train_loader = LinkNeighborLoader(
    data=train_data, 
    num_neighbors=[10],
    edge_label_index= (("recipient", "match", "donor"), edge_label_index),
    edge_label = edge_label,
    batch_size=16, 
    shuffle=True,
)

# Inspect a sample:
sampled_data = next(iter(train_loader))

print(sampled_data["recipient", "match", "donor"].edge_label_index) # ==>tensor([[6, 4, 5, 7, 0, 1, 2, 3], [6, 4, 5, 7, 0, 1, 2, 3]])
print(sampled_data["recipient", "match", "donor"].edge_label) # ==> tensor([0, 0, 1, 0, 0, 1, 1, 0])

[rusty1s]

Please note that the mini-batch assigns new node IDs based on the sampled subgraph. As such, it is not expected that your edge_label_index matches with the original edge_label_index. The remapping can be performed via

src_n_id = sampled_data['recipient'].n_id
dst_n_id = sampled_data['donor'].n_id
row, col = sampled_data['recipient', 'match', 'donor'].edge_label_index
print(src_n_id[row])
print(dst_n_id[col])

[SheidaMajouni]

That is correct!

Thank you so much! It was a great help. I was so confused.