kdgutier
commented
2 years ago Hi @manujosephv,
-
The _sEncoder of the N-BEATSx model allows you to pass categorical features: https://github.com/Nixtla/neuralforecast/blob/main/nbs/models_nbeats__nbeats.ipynb https://github.com/Nixtla/neuralforecast/blob/main/neuralforecast/models/nbeats/nbeats.py#L38-L51
-
The ES-RNN model already has level and seasonality embeddings that work as you mentioned. 2.1. Regarding the embeddings solution, they work well until the dataset grows into millions of series and explode in memory. 2.2. Or in our experience, embeddings tend to be harder to learn with little data, in the past we have preprocessed and passed the train seasonalities with the static features (inspired by the ES-RNN) with very good success. https://github.com/Nixtla/neuralforecast/blob/main/nbs/models_esrnn__esrnn.ipynb https://github.com/Nixtla/neuralforecast/blob/main/neuralforecast/models/esrnn/esrnn.py#L200
-
We are working on a multivariate extension to N-HiTs with embeddings. This is still research work.
Hope that the existing categorical options of neuralforecast models help you.
Thanks for your comment.
rcslight
Is your feature request related to a problem? Please describe. Many real life datasets have categorical features, some of them meta data about a time series(static) and some of them varying with time(exogenous). Currently the only way to include such variables is to do some preprocessing and encode those variables as numbers
Describe the solution you'd like An option to specify these features(static or dynamic) are categorical and then use an Embedding Layer to replace them with learned embedding while training.
Describe alternatives you've considered Alternative solution would be to use separate category encoders and convert these features to numbers prior to training.
Additional context This is what GluonTS has implemented.