For link prediction inference, how can I score every pair of nodes from new unseen graphs with no positive edges?

[ogggcar]

Mmm it wasn’t my function its dgl message passing function.
Function is:

graph.apply_edges(fn.u_dot_v('h')

[mufeili]

It’s the same as torch.dot. See the example below.

import dgl
import dgl.function as fn
import torch

g = dgl.graph(([0, 1], [1, 2]))
g.ndata['h'] = torch.tensor([[1., 2.], [3., 4.], [5., 6.]])
g.apply_edges(fn.u_dot_v('h', 'h', 'dot_out'))
print(g.edata['dot_out'])
# tensor([[11.],
#         [39.]])

print(torch.dot(g.ndata['h'][0, :], g.ndata['h'][1, :]))
# tensor(11.)
print(torch.dot(g.ndata['h'][1, :], g.ndata['h'][2, :]))
# tensor(39.)

[ogggcar]

Thanks, I now understand it. Then must be something else. Do you see something wrong with the following code? Maybe Im using different embeddings for each case?

feats = grafo.ndata['Feats']
updated_feats = model.sage(grafo, {'ent': feats})
updated_feats = updated_feats.values()
updated_feats = list(updated_feats)[0]
embeddings = model.pred.etype_project[('ent', 'link3', 'ent')](updated_feats)
pred = model.pred(grafo, {'ent': embeddings}, ('ent', 'link3', 'ent'))
print(pred)

This gives me the pos_score for link3 edges, and look like this:

tensor([[213.6073],
        [ 15.9547],
        [ 25.2864],
        [ 16.5195],
        [  9.9416], grad_fn=<GSDDMMBackward>)

This is used during training. However, for prediction I do the following. I calculate the same dot product between link3 edges with the following:

src, dst = grafo.edges(etype=('ent', 'link3', 'ent'))
for i in range(len(src)):
  print(torch.dot(embeddings[src[i], :], embeddings[dst[i], :]))

I get:

tensor(3274.8743, grad_fn=<DotBackward>)
tensor(240.6167, grad_fn=<DotBackward>)
tensor(412.9496, grad_fn=<DotBackward>)
tensor(238.0706, grad_fn=<DotBackward>)
tensor(128.2873, grad_fn=<DotBackward>)

Should these dot values be the same? I really dont see where I could be using the wrong embeddings…

[mufeili]

Perhaps model.pred also called model.pred.etype_project before computing the dot product.

[ogggcar]

That was the problem, thanks!