Explanation for the automatic model conversion from homogeneous to heterogeneous

[enter802]

Could someone please explain figure for the automatic conversion of a homogeneous model to heterogeneous model provided here? Although I have implemented the process successfully, I cannot make sense of the figure. Will really appreciate if someone could shed some light.

[rusty1s]

Every GNN layer is duplicated for each edge type, and every node-level operation is duplicated for every node type. Edge-type specific embeddings from a GNN layer are then aggregated to each destination node type and form the final GNN embedding of the node type.

Imagine you have 2 node types (B, C) pointing to node type A. The pseudo code then becomes:

out_A = conv_B_A((x_B, x_A), edge_index_B_A) + conv_C_A((x_C, x_A), edge_index_C_A)
out_A = relu(out_A)

[enter802]

Sorry for the very late response but I was in ER for few days due to some health condition.

I understand what you have said. But I am having trouble connecting it to the figure. If you don't mind, could you please kindly provide a pseudo code for the figure?