Use of graph pooling for node classification

[torjusn]

The usage file and another paper I found mentions the possibility of using the architecture for node classification. I wondered how the signal classification example differs from node classification with multi-feature nodes V=(number of nodes, node features) and A = (number of nodes, number of nodes), specifically:

1. For node classification I'd assume we need the final layer to output: (Batch size=1, Number of classes, Nodes) to get a prediction per class just set the last fully connected layer=number of classes, but how can we get a prediction for each node when the node pooling decreases number of nodes so the final layers number of nodes is V/(pool_size_1*pool_size_2).
2. The Notebook example expands input since features=1, so first pass to gconv is N, M, F=1, i.e. graph signals in a batch, number of nodes, and features=1. For this case would rank still be=2 since data only has number of nodes in batch and features, and should it be expanded? It doesn't seem like there should be any M if not dealing with graph signals.

Thank you for the great article!