Good to Batch node features running message passing on a static graph?

[songyesog2000]

Hi there, I am looking for an example to train GCNN with fixed graph structure while the information on the nodes is dynamic.
I wonder can I batchlize the data with one extra dimension (like general nn batching) and run it on the known fixed graph (message passing process ), rather than transforming multiple graph into a large sparse matrix. Is that reasonable and possible to do so? If there is some example training like this, I would be happy to read those codes.

Essentially, I wondering whether the default batch method (SpMM ish) will reduce the model bias for the graph structure, otherwise, I can just try to batch the data as default.

Thank you so much!

[rusty1s]

For some GNNs, it is possible to use [batch_size, num_nodes, num_features] as node feature matrix as input, e.g., GCNConv, SAGEConv, ..., while using a fixed edge_index across the batch dimension. Hope this helps :)

[songyesog2000]

Will have a look of it thanks!