Problem with edge classification on graph

[JHZhou123]

Hi Matthias,

I am working on a problem that the dataset is different graphs and the label is one of the existing edge in each graph, so I try to use GNN to train the graphs and use

out=out[row]+out[col]
F.log_softmax(out, dim=1)

to find the edge with the highest probability, but in this case global pooling is not used, so how to use mini-batching ?

I also find out if the model predict the label (one edge of the graph) directly, the result is good. But if I train the graph, upgrade node attribute and use the mlp(node1+node2) to find out the edge with the highest probability, the model is not working and the result is poor, maybe this method does not make sense. Can you help me?

thank you!

[rusty1s]

It's tricky to implement that in a mini-batch fashion, as you need to ensure that your positive link exists in the mini-batch. I guess you could perform subgraph sampling, where you ensure that your positive link exists in the subgraph, and then proceed with your pipeline as usual.

[rusty1s]

I'm not really sure what that means TBH. Your training method should look similar to:

out = out[row] + out[col]
prob = F.softmax(out, dim=0)
loss = F.cross_entropy(prob[edge_idx], torch.tensor([1])

[JHZhou123]

I use
out = out[row] + out[col]
out is [edge_number,1]
then

log_prob = F.log_softmax(out, dim=0)
loss=F.nll_loss(out.T,edge_index)

but the performance is poor, is this a good way to do it?

[rusty1s]

No, that does not make much sense from my point of view. If you want to find the most important edge from a set of edges, then you should do something like I mentioned above (where edge_idx represents the index of the ground-truth edge).

[JHZhou123]

Thank you, but do you mean F.binary_cross_entropy_with_logits instead of F.cross_entropy in the last line?

[rusty1s]

Yeah, you need binary_cross_entropy. The softmax needs to be applied beforehand.