Implementing skip connections in GCNs

[codeastra2]

Skip connections are a useful feature to overcome issues of over-squashing:

1. https://minyoungg.github.io/MIT-deeplearning-blogs/2021/12/09/oversquashing-in-gnns/
2. https://arxiv.org/pdf/1806.03536.pdf

When aggregating some of the information from the nodes we skip some of these nodes, could we add a argument in GCN(in_f, out_f, skip=True) which would do this?

[rusty1s]

I think we already provide this option via the jk flag, e.g., in GCN or SAGEConv. There also exists a separate JumpingKnowledge module.

[codeastra2]

@rusty1s thanks for your reply, I searched the documentation of

1. https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#convolutional-layers
2. https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/nn/conv/message_passing.html
   and even GraphSage but did not find "jk" operator, could you please let me know the same.

[rusty1s]

See https://pytorch-geometric.readthedocs.io/en/latest/generated/torch_geometric.nn.models.GraphSAGE.html#torch_geometric.nn.models.GraphSAGE