Obtaining Embeddings in PYG.KGE and inverting to recover original input data

[blainehill2001]

Hello,

As I understand in pytorch geometric models such as RotatE or TransE we can obtain an embedded triplet for some input triplet once we've trained our model via node_emb or node_emb_im in the case of RotatE. How can we recover the original triplet if we have the embedded triplet? If we can't guarantee the reverse process since we are mapping to a latent space with fewer dimensions than the input dimensions, how can we get the most likely triplet that matches our embedded triplet? Looking at the source code for torch_geometric.nn.kge there doesn't seem to be a decode() or anything like there is in this example.

To ground the question in a more detailed manner:
Let's say I have a trained embedding model and wish to do link prediction or some reasoning task. I take my (head_node_index, unknown_index, tail_node_index) and embed it using my model (let's say RotatE).

I now have six objects: model.node_emb(head_node_index), model.node_emb_im(head_node_index), model.node_emb(tail_node_index), model.node_emb_im(tail_node_index), model.rel_emb(unknown_index), model.rel_emb_im(unknown_index)

Then I plug these objects as input to some other model which yields an answer for the link prediction in the embedding space. I now wish to map my six objects to the input dimensions so I can have a meaningful answer.

[rusty1s]

Sorry for super late reply :( Is there a point in a decode function if you already node head_node_index, tail_node_index and unknown_index (since you receive the embedding from it)? Why do you need to recover these indices from the embedding itself? Otherwise, I would assume you can simply search for the index in node_emb that matches with model.node_emb(head_node_index)?

[blainehill2001]

The goal is to obtain a predicted triplet in the input dimensions (aka natural language such as English) from a embedded triplet/sextuplet.

I think you've answered the question with your last line: if I have some sextuplet, I take each value and search for the index that yields nearest it. I originally wondered if there was any predefined function to do this (like seen in the example link).