Positive and Negative Sampling using NeighborLoader

[wilcoln]

Hello,
I am trying to train a GraphSAGE model in an unsupervised manner on the Reddit dataset. For this, I am taking inspiration from the unsupervised graphsage example to adapt the supervised Reddit example. In the former example, positive and negative sampling is achieved by appropriately extending the NeighborSampler class. However, NeighborSampler has now been deprecated, and the latter example instead makes use of NeighborLoader. Unfortunately, I couldn't find any example showing how to use/extend this latter class to perform positive and negative sampling. I've been struggling on this for a couple of days now. Any help will be warmly welcomed!

[wilcoln]

Update: I managed to train the model using the deprecated NeighborSampler.

[rusty1s]

That sounds good. Glad that you got it working.

I will also try to convert this example to make use of the NeighborLoader class. We are working on better support for this in #4026.

[daviddavo]

How did you do it? I want to do BPR and I also need positive/negative sampling with NeighborLoader, but I don't know how.

[rusty1s]

Take a look at LinkNeighborLoader :)

[daviddavo]

I want something between NeighborLoader and LinkNeighborLoader: Something that can return a triplet (like LinkNeighborLoader, just what I want, yay!), but iterates through every (source) node just once, instead of returning a sample for every link.

[rusty1s]

What about initializing two NeighborLoader instances (the second with random seed nodes), and iterating through them via

for batch1, batch2 in zip(loader1, loader2):
     pass

[daviddavo]

The seed nodes need to be the same. I need negative samples for each source node in a bipartite graph.

[rusty1s]

In this case, you can specify a sampler argument to both loaders to guarantee that they share the same seed nodes.