Question about the sampling code about ladies

[acbull]

Sorry which line are you referring to?

We don't assume the lowest layer contains all the nodes, as we are conducting layer-wise sampling. For each GNN layer, the source nodes and target nodes could be different. (which is stored as a specific adjacency matrix in https://github.com/acbull/LADIES/blob/303036e8649eaf269eb5adb87f6b57c168c39276/pytorch_ladies.py#L117-L120)

What you refer to is more like graph-wise sampling, by sampling a subgraph and then apply GNN on the whole sampled graph, which is different from our setting.

[chocolate9624]

Thanks for your quick response!

https://github.com/acbull/LADIES/blob/c10b5260ebcedf4efa7a175ec60ce6229a2a444e/pytorch_ladies.py#L157
This line returns the "previous_nodes". The 'previous_nodes' is the nodes from the lowest layer.

https://github.com/acbull/LADIES/blob/c10b5260ebcedf4efa7a175ec60ce6229a2a444e/pytorch_ladies.py#L241-L246
In these lines, the 'previous_nodes' is renamed as 'input_nodes'. And these nodes ('input_nodes') are used to slice node features for the sampled subgraph.

I think the 'previous_nodes' from the lowest layer may not contain all the nodes in the sampled subgraph. The node features for some of the nodes are not used.

Thanks!