graphpku / neuralcommonneighbor Goto Github PK

Great thanks for opening the code, and sharing the new idea for link prediction task.

I tried to get the specific code for implementing neural common neighbor operation. But it seems not easy to target the right part. In the paper, neural common neighbor operator is based on the GNN representations of common neighbors. However, the train function in NeighborOverlap.py looks like missing the code for obtaining common neighbor representations before the predictor layer. Maybe I miss something important. Could you please give some hints to help me better understand the code?

Best

Hello,
Xiyuan, @Xi-yuanWang

Thanks for sharing the code.
I found that the code and the paper don't seem to match. In the paper, Equation (16) uses the concatenation of two nodes' representation and their common neighbors' representation. But in the code of CNLinkPredictor class, it seems summation is used:

cn = adjoverlap(adj, adj, tar_ei, filled1, cnsampledeg=self.cndeg)
xcns = [spmm_add(cn, x)]
xij = self.xijlin(xi * xj)

xs = torch.cat(
    [self.lin(self.xcnlin(xcn) * self.beta + xij) for xcn in xcns],
    dim=-1)

Besides, torch.cat() in this code seems useless since there is only one tensor in xcns. Please correct me if I have misunderstood.

Best wishes,
Lei

Thank you for your excellent work! I want to use AUC metrics on my own dataset for evaluation. But I can't find the logits of the predictor output. Could you please provides some details on how to use AUC for evaluation?

Hello,

I wanted to run NCNC on my own dataset. Can you guide me as to how to do so?

Thank you!

Hi,

I am trying to apply NCN/NCNC to other graphs. In the README, it seems there are a lot of hyperparameters to tweak with. Are there any general suggestions about where to start the hyperparameter tuning?

Thanks,

Dear authors,

Thank you for sharing the code to your paper on NCN/NCNC.

I had a query regarding the use of use_valedges_as_input usage in your model in comparison to BUDDY/ELPH (Table 3 results). If I am not mistaken, the authors of BUDDY/ELPH indicate in their appendix that validation edges are also consumed as part of training message passing edges for both Planetoid and ogbl-collab datasets at the testing time. I am including a partial sentence indicating this:

"... but for the Planetoid and ogbl-collab datasets, the message passing edges at test time are the union of the training message passing edges and the validation supervision edges"

Taken from: Chamberlain, Benjamin Paul, et al. "Graph Neural Networks for Link Prediction with Subgraph Sketching." arXiv preprint arXiv:2209.15486 (2022).

Is there any reason NCN/NCNC does not use use_valedges_as_input for the Planetoid datasets runs?

Warm regards,
Paul