Regarding use of RGCN

[urmi22]

I have attached pictorial representaion of a model here. In that model I want to use RGCN-CONV, where input is a graph (graph consist of two types of nodes : documents and concepts; and three types of relational edges : document to document, document to concept and concept to concept) and output is a prerequisite prection between two concepts. I just need to know how I can implement this entire model using RGCN.
each node type and edges have features.

paper link. https://aclanthology.org/2021.naacl-main.164.pdf

[rusty1s]

I suggest to read up how PyG handles heterogeneous data here. Let me know if you have any questions.

[urmi22]

Thanks @rusty1s for your reply!

I'll try to explain the problem I facing for now. My heterogeneous graph metedata looks as follows-

HeteroData(
  concept={ x=[411, 50] },
  document={ x=[599, 50] },
  (document, contains, concept)={
    edge_index=[2, 1237],
    edge_attr=[1237, 1]
  },
  (document, related-to, document)={
    edge_index=[2, 358801],
    edge_attr=[358801, 1]
  },
  (concept, similar-to, concept)={
    edge_index=[2, 3118],
    edge_attr=[3118, 1],
    y=[3118]
  }
)

This I have created following the tutorial you have pointed out. I have defined the RGCN model as follows to extract node representation. The code looks like this -

 class Model(nn.Module):
	def __init__(self):
		super(Model,self).__init__()
		self.rgcn_layer = RGCNConv(in_channels=50, out_channels=120, num_relations=3, num_bases=3)
		
	def forward(self, x, edge_index, edge_types):
		x = self.rgcn_layer(x, edge_index, edge_types)
		return x
model = to_hetero(Model(), data.metadata(), aggr='sum')

However, if I call the forward method, I get the following error:

>> output = model(data.x_dict, data.edge_index_dict, data.edge_types)
*** AttributeError: 'List' object has no attribute 'get'

Can you please point me out the mistake I am making?

[rusty1s]

I see. I don't think it makes sense to combine both to_hetero and RGCNConv. You either need to convert your heterogeneous graph into a "typed" homogeneous one via data.to_homgeneous() from where you can then make use of RGCNConv. Alternatively, you define a regular homegeneous GNN (e.g., via SAGEConv), and then apply to_hetero on top of this model.

[urmi22]

@rusty1s Thanks! I also had this similar doubt. I'll try what you have suggested and let you know if I face any other issues.