Distributed graph classification on multiple machines

[gamdwk]

I am a beginner of GNN. I have learned some node classification content so far, and found that there is no tutorial on graph classification on multiple machines on the official website. My dataset has a large number of graphs, so I hope to achieve this. Is multi-machine graph classification feasible? If feasible, how to split the dataset? Is it to split multiple graphs to different machines, or to split each graph? Can the partition and sampling method of node classification be used in graph classification?

[frozenbugs]

In graph classification tasks, the size of each graph is usually small and independent, so you can split your dataset across multiple machines naturally by random put a subset of graphs on each machine. For the training on each machine you can follow this tutorial.
https://docs.dgl.ai/en/latest/guide/training-graph.html

[gamdwk]

Thanks. So the distributed sampling method is meaningless in graph classification, right? We no longer need to obtain nodes and features remotely.

[mufeili]

At least it’s very different from node classification. It will be similar to distributed image classification.