2
$\begingroup$

If you are building a DNN, say, with two layers, and you want to use embeddings as one of your feature inputs, what's the best way to input the embedding?

I'm trying to understand if I should break the embeddings up so that every array value in the embedding becomes its own input feature to the model or whether the embedding should be kept in array form.

I've been following AirBnB's model for inspiration.

I'm trying to predict a binary classification in the final layer.

$\endgroup$
1
$\begingroup$

Breaking the array you lose distributed representation-point of word embedding.

There is information in other dimensions (words) that are distributed in the word-embedding representation, breaking them up makes no sense.

$\endgroup$
1
  • $\begingroup$ I assume those would be re-learned within the network itself to some extent.. but I see and that makes sense. $\endgroup$ – Learning stats by example Jan 2 '20 at 23:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.