[ENH] interface probabilistic regressors from `ngboost` package

[sktime]

[ENH] interface probabilistic regressors from `ngboost` package

[fkiraly]

well, if you want to use skpro for probabilistic predictions, the package itself needs an update - it's a bit of a larger project than just adding a method. If you're interested to do that, we should set up a call - also (I believe you are part of the mentoring programme?) discuss with your mentor.

[fkiraly]

actually yes!

We've been re-working the probabilistic forecasting interface:
sktime/sktime#4359

This will enable using probabilistic supervised learners in compositors like make_reduction much more easily.

Want to help work on this, @drackham ? It's a bit of an engineering project, but there's a step-by-step roadmap.

Would be much appreciated!

We'll probably move this a bit over the Easter holdays where the volunteer contributors tend to have more time.

What would also be helpful is testing the probabilistic forecasting interface and reporting your experiences or any design suggestions (in sktime/sktime#4359). Will be released experimental in 0.17.0 and full in 0.18.0.

[fkiraly]

(will continue on sktime/sktime#4359 for architecture discussion)

[ShreeshaM07]

I'll try to start working on adding an interface of ngboost to skpro soon after going through its docs.

[fkiraly]

Interesting!
NGboost is a probabilistic supervised prediction algorithm (for tabular data), not a forecaster - so we would first have to build a full probabilistic interface. That would be a great thing to have for sktime.

Now about something that's partly funny and partly not funny...
"NGboost" is basically identical to the probabilistic boosting algorithm I proposed (earlier) in section 6.4 of my 2018 paper https://arxiv.org/abs/1801.00753
The probabilistic prediction interface of the NGboost python package closely follows principles in the sktime companion package skpro (scikit-learn like probabilistic prediction), or the R package mlr3proba.

I'm reasonably certain that Ng et al know of both the methodological paper and the software interface designs, but still don't cite them... that's not very nice.

Anyway, we should develop skpro into the probabilistic scikit-learn that it was meant to be, but unfortunately it's currently without maintainer. Might that be something you would be interested in?

[satya-pattnaik]

Yes, 6.4 of the above mentioned paper deals with a similar concept.
That sounds good to me, I can use Skpro to build it. The question is how to build that probabilistic interface, if I can get some starting point I can carry it from there, as I was planning to use make_reduction(Which I assume now is not the correct way???)

[satya-pattnaik]

Yes, will discuss about this with my mentor. Thanks

[drackham]

I'm very interested in the capability discussed here. It doesn't appear like there has been any progress here. Would you correct me if that is in fact not the case? Thank you!

[fkiraly]

speaking of which, @frthjf, are you still around?

I would like to move the probabilistic interface into skpro within the next year or so and use it as an import.

That would be step number 7 or 8 in sktime/sktime#4359 (not there yet, but see context above).

[drackham]

@fkiraly thank you!

I discovered sktime recently, and these types of models are not really in my core areas of competency, so I'd likely be unable to contribute effectively. That said, I'll take a look at the contribution documentation and see if my apprehension is unwarranted.

I'll also take a look at the probabilistic forecasting interface and report back. Thanks again!

[frthjf]

@fkiraly I am still around :-)

Just to clarify, do I understand correctly that the plan is to resurrect the skpro package which in turn becomes an optional dependency of sktime? I could certainly help with updating the skpro code.

[fkiraly]

@fkiraly I am still around :-)

Nice to hear of you again! Let's catch up, discord perhaps?

Just to clarify, do I understand correctly that the plan is to resurrect the skpro package which in turn becomes an optional dependency of sktime? I could certainly help with updating the skpro code.

Yes!

For now, I've been working in the sktime/proba module. Have a look and let me now what you think!

The design is a mix of pandas, sklearn (base interface), and tensorflow-probability (parameter broadcast).

I'd like to move it out to skpro, and make skpro a core (not optional) dependency of sktime. Both sktime and skpro would eventually depend on skbase which has the base class framework.

[fkiraly]

For comments, the topical issue is here, @frthjf:
sktime/sktime#4359

(this issue is about a specific probabilistic forecaster)

[frthjf]

I see, in that case, why move this out of sktime into an independent skpro package if it's required back in anyway? Wouldn't it be easier to port the skpro features into sktime instead? It seems to me that the sktime package already has all the CI and package infrastructure we would have to recreate for a skpro re-release.

[fkiraly]

The skpro package is now sufficiently mature to accommodate an interface to ngboost - moved the issue therefore to skpro.

[fkiraly]

FYI @drackham, @satya-pattnaik, @frthjf - I have updated the issue with instructions on "how-to". Together with the skpro machinery and the existing integration to sktime, this should now be pretty straightforward.

Would be great if one of you would like to implement this, I'm happy to advise!

Also FYI @Alex-JG3, @Ram0nB, in case one of you is interested, this intersects with your previous contribution topics.

[KiwiAthlete]

Have you considered XGBoostLSS and LightGBMLSS as well? Both offer great flexibility and are based on the two most commonly used tabular data boosting machines. Yet, there is no sklearn API available, but has a PR on this.

Shall I open a new issue for this?

[fkiraly]

Excellent suggestion, @KiwiAthlete - opened an issue here: #135

[fkiraly]

Great! I notice I linked the wrong issue, fixed the link.
I'd recommend you continue discussing in the issue #135 how the interface would look like.