Link prediction using graph without attribute tensor

Since my graph has no node attributes, I initialize the a 4-dimensional tensor which is [1,1,1,1] to every node. I am wondering if this would cause any side effects when training model?

If your graph has no node attributes and you initialize the node features with the same vector, then your model will not be able to differentiate between isomorphic nodes, which is usually critical in link prediction.

Say that your graph looks like this:
Then with your GNN model like above you will not be able to differentiate node 1 and node 4, so you won’t be able to differentiate between the pair (1, 2) and (4, 2).

There are several papers discussing this issue, including [1910.00452] On the Equivalence between Positional Node Embeddings and Structural Graph Representations that discusses the necessity of such differentiation in theory, and [2010.16103] Revisiting Graph Neural Networks for Link Prediction giving a practical solution (you can also check out the SEAL example in our repository).

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.