can you provide a classification example?

[carrielui]

I would be nice to have an example of classification problem.

[avati]

Hi @carrielui, thank you for filing the request. Currently our focus has been on regression problems (since the project started under a research problem that involved probabilistic regression), though the framework could trivially be extended to classification.

What needs to be done is: implement a new ngboost.distn.Bernoulli or ngboost.distn.Categorical for binary and multi-class respectively, and classification should just work without any other changes.

Separately, we also welcome community contributions, and happy to help / guide someone to do this if it is required sooner than me getting to it.

[wptmdoorn]

I actually wrote a ngboost.distn.Bernoulli class last night but I have to clean/adjust and test it. But I am happy to share it once I am done. I am currently busy with some blog post concerning the use of NGBoost in medical regression problems, but once that's covered I will work on the Bernoulli class. I expect to finish it early next week.

[avati]

Great, will look forward for the PR @wptmdoorn!

[romanvoyt]

Hello! I want to use ngboost in bankcruptcy prediction tasks. So, as you said above we need to create ngboost.distn.Bernoulli or ngboost.distn.Categorical. Is it important to have both of them?

I would like to help develop this anyway.

[wptmdoorn]

@romanvoyt I've been busy quite a lot but I expect to upload a first version of the Bernoulli class tonight (in about ~7-9 hours). I would be more than grateful if you would jump in on the project :) I guess ngboost.distns.Bernoulli and Categorical share a lot of common theory and implementation so I think it shouldn't hurt to implement them both.

[alejandroschuler]

@wptmdoorn hell yeah! I'm excited to see this functionality come to life. Would you mind also adding a classification example to the readme to mirror the regression example?

Another thing that would help us would be to have a "how to write a distribution for ngboost" tutorial, written in plain english, so that other folks can add more distributions in the future (e.g. beta, poisson). As someone who has just gone through the process yourself, but understands the perspective of a user, would you be interested in having a stab at that?

[wptmdoorn]

@alejandroschuler I would be willing to do this, but I have to admit I am fairly new to open-source projects and also the amount of math/statistics involved in this project is a lot for me, personally. Once the Bernoulli class is implemented I could give it a go (but there also might be better persons to write such a guide!)

@romanvoyt I just submitted an issue #17, giving it a first try, any help would be really welcome.

[romanvoyt]

@wptmdoorn Great. Let me check this one later, I hope I can help with it.

[tonyduan]

Closing this issue out since we now have a classification example.