Source: Deep Learning on Medium
I have a gut feeling that the embedding+dot prod+sigmoid is very similar to matrix factorization in recommendation systems?
In our case we could make a matrix of Books x Wikilinks and then factorize the matrix to have a Book-Features (Book x Embedding dim) and (Embedding dim x Wikilink) matrix.
the Book x Embedding dim matrix then can be used for book embeddings as well I guess.
So my question is What advantage does DL based embedding have over the matrix factorization?