Can the default implementation of GCN in DGL tutorial handle heterogeneous graph

[baishakh]

I was looking into this tutorial:
https://docs.dgl.ai/en/0.4.x/tutorials/models/1_gnn/1_gcn.html

It appears this GCN implementation can only handle a homogenous graph. Is this understanding correct?

Secondly, is there a GCN implementation available with DGL that can handle heterogeneous graphs?

[minjie]

Hi, the GCN-for-heterograph is actually relational graph convolutional network (RGCN). We have two implementations:

• dgl/examples/pytorch/rgcn at master · dmlc/dgl · GitHub when the node and types are stored as node and edge features.
• dgl/examples/pytorch/rgcn-hetero at master · dmlc/dgl · GitHub when the graph is in native DGL heterograph (i.e., created by dgl.heterograph).

[baishakh]

Thanks @minjie.

Can you please elaborate what’s the difference between this two implementations Which one would you recommend using?

[minjie]

Generally, if your graph has many relations and node types (e.g., >100), I will recommend the first implementation. The second one will be more efficient if your graph has few relations and node types.

[baishakh]

I would use this for graph classification and it would have quite large. So sounds like I should use the first implementation.

Thanks, @minjie for the feedback.