Doubt negative sampling for link prediction

[yorard123]

Hi,

I'm trying to do a negative sampling for a link prediction problem. My problem has two kinds of nodes, type1 and type2. The edges only can be between type1 and type2. I have been checking the documentation and I found that the current methods for batched negative sampling does not support this use case. Is there anyway in pytorch geometric to solve this issue? I was thinking to do it manually creating manually an all to all edge_index vector, and pick randomly certain number of edges, however, as I did it was very slow. I would appreciate any hint on this issue.

EDIT: Another requirement that I have is that I need that each node of type1 has the same number of edges.

[rusty1s]

Thanks for pointing this out. I made some changes in master, and you can now perform bipartite batch-wise negative sampling as well, see here. This is achieved via a tuple of batch vectors as input, i.e.:

batched_negative_sampling(edge_index, (src_batch, dst_batch))

I will craft a new release soon.

[yorard123]

Hi,

Thanks! I have one doubt related to the src_batch and dst_batch. At this moment what I have is a data structure that contains all the nodes in there, and to know whicch are of type 1 or type 2 I have a mask that it helps me to know that. For instance:
data.x = [x, x, x, y, y, y, x, x, x, y, y] data.batch = [0,0,0,0,0,0,1,1,1,1,1] data.mask_1 = [1,1,1,0,0,0,1,1,1,0,0]

Assume that x is type 1 and y type 2. My question is, what is expected to be the src_batch and the dst_batch?

Thanks!

[rusty1s]

src_batch and dst_batch refer to mini-batching with HeteroData objects, which is currently the recommended way to operate on heterogeneous graphs. That is, different node types are stored in distinct matrices, and are connected via bipartite graphs. For your use-case, it's probably best to simply over-sample the amount of negative edges, and then only keep the ones that connect different types:

neg_edge_index = batched_negative_sampling(edge_index, batch, ...)
mask = data.mask_1[edge_index[0]] & ~(data.mask_1[edge_index[1]])
neg_edge_index = neg_edge_index[:, mask]

[yorard123]

Thanks!