Does dataloader destroy graph structure using mini-batch?

[jiaqian]

Hi, I create a dataloader (e.g. NeighborLoader) by a big graph G(composed of some subgraphs), I am wondering does batch likely destroy the graph structures (for instance breaking some edges) with small batch size since I did not supply a list of subgraphs to dataloader? Any input is welcome, thanks.

[rusty1s]

Can you clarify? I am not totally sure I understand, sorry. If you apply the neighbor loader on a big supergraph containing many isolated subgraphs, these subgraphs will still be disconnected after sampling. Importantly, not all subgraphs will be complete, depending on the number of hops and neighbors you sample.