Pytorch_geometric: Unsupervised GraphSAGE

[yuanx749]

❓ Questions & Help

Hi, I would like to implement unsupervised GraphSAGE in the GraphSAGE paper. Is there a function in PyG for sampling the negative node indices? torch_geometric.utils.negative_sampling is used to sample edges but not nodes.

[rusty1s]

Hi and thanks for your interest! We do not have an example for this yet, but you can follow this issue to track its progress.

[yuanx749]

@rusty1s I have implemented unsupervised GraphSAGE, using torch_cluster.randow_walk to sample neighbors and NeighborSamplerto generate mini-batch, and another NeighborSampler to randomly shuffle the nodes for negative sampling.

Could you have a look at the code below? It would be great that it could be added to the examples.
https://gist.github.com/yuanx749/de835f92f4e44a9cf825a5638968e7e5

[arunavsk]

Hey @yuanx749 - I have a simpler version of this https://gist.github.com/arunavsk/7a4091ccddcbfa6eb31c35c5ce7fe462

[rusty1s]

Both versions look fine to me. We can definitely integrate it into the PyG examples. Feel free to send a PR.