When to perform node/edge graph feature extraction?

[AlessioBoetti]

Hi! I have a CSV file which can be converted into a PyG graph data object for an edge classification task. Before doing that, I thought of adding some features using NetworkX library.

However, since after creating the graph I’m going to split it into train/val/test, then using a dataloader, I am unsure if the graph-based preprocessing should be done before the split, right after the split (before creating the dataloader), or inside/after the dataloader construction.

If the answer is inside/after the dataloader construction, does this mean that I have to implement a custom transform for every feature I want to add, that basically transforms the graph to a networkx graph, extracts the feature, then reconverts the graph to pyg?
What’s the best way to proceed?

Thank you!