Dynamic Categorical Features

[andreimardale]

Hello and thank you very much for this nice library!

I am using the DeepAR model in order to predict daily cumulative sales for a bunch of products. One of my exogenous features would be the status of a product (some products can be inactive, thus the cumulative sales of those product will not increase as long as the product is inactive). There are 9 such statuses.

I would encode them as dynamic real features, but it looks more like a categorical feature to me.

Therefore, I believe that a good solution would be to use the dynamic categorical features. However, in the DeepAR class, these are not implemented. I watched the issues regarding this subject and someone was suggesting to use Embeddings and then to append these to the dynamic real features.

Could you please help me with an example of how such functionality could be achieved? I checked the code, but I am not really at ease with mxnet. :)

Thank you very much!