Differences between `idgnn_link.py` and `gnn_link.py`

[quang-truong]

Hi,

I have tried to differentiate between the examples idgnn_link.py and gnn_link.py. There are few notables differences:

• idgnn_link.py uses NeighborLoaders for train, validation, and test sets. gnn_link.py uses LinkNeighborLoader for train, and a pair of source and destination NeighborLoaders for each validation and test set.
• LinkNeighborLoader has the share_same_time argument.

If I understand correctly, idgnn_link.py samples nodes, and the model learns the embeddings of nodes in the destination tables. Meanwhile, for gnn_link.py, there are sampled positive and negative destination nodes. In this file, the model minimizes the scores, which is calculated by element-wise product of source and destination embeddings.

Please correct me if I understand it wrongly. My main questions are:

1. Which setting is officially adopted for Link Prediction in the RelBench paper?
2. What is the intuition behind share_same_time argument for LinkNeighborLoader? And why does it affect how negative scores are calculated? Below is a snippet from gnn_link.py.

   pos_score = torch.sum(x_src * x_pos_dst, dim=1)
   if args.share_same_time:
   # [batch_size, batch_size]
   neg_score = x_src @ x_neg_dst.t()
   # [batch_size, 1]
   pos_score = pos_score.view(-1, 1)
   else:
   # [batch_size, ]
   neg_score = torch.sum(x_src * x_neg_dst, dim=1)

3. I could not find details about samplers for each dataset in the RelBench paper. I realized there are other arguments that are not covered in the paper, such as temporal_strategy, subgraph_type, etc.