Limitation on training GCNs for large graphs

Questions
#1

If the task is to learn node embeddings for node classification, what is the maximum size of the graph that I can load into dgl? I’m trying to apply it within hardware circuit graphs, which are represented by millions of nodes.
How does dgl handle large graphs? Or is there actually a limitation on training GCNs for large graphs? Any references to papers/code will be much appreciated.

Thanks,

#2

Hi,

Millions of graph is supported and we have customers using this in their recommendation system. For large graphs, sampling is needed. You can find the example at https://github.com/dmlc/dgl/tree/master/examples/pytorch/sampling

1 Like