Heterograph link prediction with nodes have different dimensions and no reverse edges

[tsuresh83]

Hi,
I am just getting started on the dgl library - thank you for the wonderful library. I am following the hetero graph link prediction example from here on my own dataset. It has 2 node types drug and disease and one edge type ‘treats’. The drug nodes have 500 dimensional feature vector and the disease nodes have a 300 dimensional feature vector. Also the ‘treat’ relation is directional going from drug to disease. I understand that this dataset is different from the example heterograph here in 2 important ways

1. the feature dimensions are different for the 2 node types
2. there are no reverse edges going from disease to drug.
   I can’t get the example code on the link to work with these 2 differences. Can someone please help me with the required modifications?
   Thanks,
   Suresh

[BarclayII]

It seems that you were using 0.8.x documentation. You could see this example: dgl/examples/core/rgcn at master · dmlc/dgl · GitHub

As per the reverse edges, I would recommend you still add the reverse edges (with a different edge type) even if the original graph does not contain them. This is just to allow the messages from a node to flow backwards along the edges.