Use CatBoost MultiQuantile regression

[swertz]

Is your feature request related to a current problem? Please describe.

CatBoost has the ability to regress multiple quantiles using a single model, with the MultiQuantile loss: https://catboost.ai/docs/en/concepts/loss-functions-regression#MultiQuantile

However, it seems that feature is not used by Darts: https://github.com/unit8co/darts/blob/master/darts/models/forecasting/catboost_model.py#L352

It might make the use of multiple quantiles much faster by avoiding the training of multiple models.

Describe proposed solution

It seems relatively straightforward to use MultiQuantile instead of Quantile. I'm happy to contribute a PR for this. The questions I would like to clarify are:

• how does this interact with multi-output regression, either for multiple components, or in the case of multi_models=True with output_chunk_length>1? As far as I understand, multi-regression is only supported for RMSE loss, correct?
• should this be left exposed as an option to the user?

[dennisbader]

Hi @swertz and thanks for raising this feature request. It's true that currently all SKLearn-like models that support likelihood="quantile" do so by fitting dedicated models per quantile.

We'd be very happy about a contribution :)

Regarding your questions:

• We perform some checks here until line 1059 to determine:
  • is multi-output required? (e.g. either multivariate target and / or output_chunk_length>1 when multi_models=True)
    • if yes, does the model support multi-output natively? For multi quantile it would have to support multi-quantile for multiple columns
      • if yes, use the model as is
      • if no, wrap the model in our MultiOutputRegressor that fits a dedicated model for each target component and step in the output chunk
• yes, we should expose this as a new likelihood option. We could call it "multiquantile" or similar.

How does that sound?

[swertz]

Thanks for the feedback @dennisbader !

if yes, does the model support multi-output natively? For multi quantile it would have to support multi-quantile for multiple columns

Indeed, as far as I understand this only works for using multiple quantiles for a single regressed target, so a multivariate target or output chunk would still need to be handled through the MultiOutputRegressor.

I'll try to have a look at this :) Would you mind giving a few pointers regarding where/how to add tests?

[dennisbader]

Hi @swertz, sure :)
You can add dedicated CatBoost tests to darts/tests/models/forecasting/test_sklearn_models.py. There are already a couple of tests for multi-output available. But a dedicated test would make most sense to me.

One thing that we should also keep in mind is to make SKLearnModel.get_estimator() work with the new multi-quantile case.

And on a last note, if multi-quantile is also natively available for other models (LightGBM, XGBoost, ...) then we could add support for that as well :)