Is there a version of embedding layers for neural networks that allows for multiple class-membership features? Any frameworks that have implemented this?

E.g. imagine we are trying to predict something about a product development program and we are interested in using the feature of which company is developing the product. However, occasionally two companies co-develop a product. Let's assume I think that the effect of these two companies being involved is in some sense getting averaged out (rather than ending up something so different, because they operate completely differently when cooperating than when working alone - in which case a completely separate category might be needed).

The main thing I've come up with, so far, is to create records with all possible values of the class features and to weight them according to 1/(number of applicable classes) during training. I kind of suspect there must be something more elegant than this.

  • $\begingroup$ average the company embeddings? $\endgroup$ – shimao Dec 1 '19 at 21:31

Your Answer

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

Browse other questions tagged or ask your own question.