question about edge_index of the reverse links in a mini-batch from LinkNeighborLoader

[NadiaAlbukaeiDH]

Hi PyG Community,

i am using LinkNeighborLoader to generate mini-batches from my hetero graph (train_data) as follows:

 train_loader = LinkNeighborLoader(
    data=train_data,
    num_neighbors=[30, 20], 
    neg_sampling_ratio=1.0,
    edge_label_index=(("customer", "order", "vendor"), train_data["customer", "order", "vendor"].edge_label_index),
    edge_label=train_data["customer", "order", "vendor"].edge_label,
    weight_attr="edge_weight",
    batch_size=1024,
    shuffle=True,
)

Here is the structure of my train_data :

HeteroData(
  customer={ node_id=[596377] },
  vendor={ node_id=[13920] },
  (customer, order, vendor)={
    edge_index=[2, 1782039],
    edge_weight=[1782039],
    edge_label=[763731],
    edge_label_index=[2, 763731],
  },
  (vendor, rev_order, customer)={
    edge_index=[2, 1782039],
    edge_weight=[1782039],
  }
)

When I checked the edge_index in one batch from the LinkNeighborLoader, I found that the edge_index of the direct and reverse links (('customer', 'order', 'vendor') and ('vendor', 'rev_order', 'customer')) are not symmetrical:

sampled_data = next(iter(train_loader))
print(sampled_data[('customer', 'order', 'vendor')].edge_index)
print(sampled_data[('vendor', 'rev_order', 'customer')].edge_index)

tensor([[ 2041, 2042, 2043, ..., 94806, 94807, 42500],
[ 0, 0, 0, ..., 5805, 5805, 5805]])
tensor([[ 1816, 391, 1817, ..., 1127, 10399, 4496],
[ 0, 0, 0, ..., 43457, 43457, 43457]])

Shouldn’t they be symmetrical for correct message passing? Why doesn't LinkNeighborLoader have a parameter similar to rev_edge_types in RandomLinkSplit, which could specify that these are reverse edges and prevent them from being sampled separately?