Question: Is it valid to use ComplexDiagonalDynamicOperator and DotComparator in link prediction and IndexFlatL2 in Nearest neighbor search?

[Ishitori]

Hi there,

I am reading the doc of how to use generated embeddings at https://torchbiggraph.readthedocs.io/en/latest/downstream_tasks.html. I see an example how to do a link prediction, which uses the same operations as in training: ComplexDiagonalDynamicOperator and DotComparator. The same thing is done in Ranking example.

But once we are doing Nearest neighbor search, we just use IndexFlatL2... Correct me if I am wrong, but these things are not the same, right? They are going to produce different results. And, if so, shouldn't we use FAISS with METRIC_INNER_PRODUCT https://github.com/facebookresearch/faiss/wiki/MetricType-and-distances ?