How to get the node features for link prediction task?

[shahinghasemi]

Hello, I have a heterogeneous network this is the data I have:
types of nodes: {dr, di, t, p, s, e}
types of edges: dr-p, dr-e, dr-t, dr-s, dr-di which are undirected.
The final problem is to find the links in dr-di.

I extended the provided example with my data. I have a big problem. I only have feature of node di and don't have any features for other nodes. for now I used identity matrix for other nodes types which is of shape (N-node, N-node) which node is our node types.

This is my question: How can I get the (initial)features in a reasonable way for other node types to use them for message-passing process to get enriched embeddings for my node types? Is using identity matrix for node features makes sense?

[rusty1s]

Yes, this is the usual appraoch. You can also use a learnable torch.nn.Embedding for each node type (which is similar but saves memory). Alternatively, there exists some work that explore using node embeddings generated by a MetaPath2Vec model.

[shahinghasemi]

Thank you for your reply. How about edge-weight? What's the usual way to get the edge-weight? Can I extract logical edge-weight with the adjacency matrices that I have?

[rusty1s]

If your adjacency matrix is weighted, the edge_weight should be given to you, right? Otherwise, there is no need to pass in an edge_weight vector in the first place. Let me know if I am missing something.