Google Summer of Code: new loss functions in XGBoost

[tdhock]

I was wondering if there is any xgboost maintainers who would be interested to co-mentor a Google Summer of Code (GSOC) student with me? I am a machine learning researcher who is interested in using xgboost for some standard loss/objective functions which are not included in xgboost. I am also an expert R package developer and admin/mentor for R in GSOC so we could probably get the student funded under R-GSOC, https://github.com/rstats-gsoc/gsoc2019/wiki/table%20of%20proposed%20coding%20projects

However I am not an expert about xgboost internals, so it would be great if somebody who is could co-mentor. Any takers?

For (left, right, and interval) censored outputs, AFT (https://en.wikipedia.org/wiki/Accelerated_failure_time_model) losses:

• Gaussian
• Logistic For count data with an upper bound, Binomial loss = negative log likelihood of the https://en.wikipedia.org/wiki/Binomial_distribution

I found some related issues: https://github.com/dmlc/xgboost/issues/749 https://github.com/dmlc/xgboost/issues/513 https://github.com/dmlc/xgboost/issues/326

[RAMitchell]

Don't have time myself but great idea.

[hcho3]

So you don’t have to be a Google employee to be a mentor in GSOC? I am interested

[tdhock]

no you don't have to be a google employee!

It would be great to co-mentor with you @hcho3

I have started writing a wiki page for the project here https://github.com/rstats-gsoc/gsoc2019/wiki/xgboost-loss-functions

can you please add some info to the project and tests sections? The tests should be some simple to complex tasks that the student can do prior to GSOC in order to demonstrate to us mentors that he would be capable of doing the project.

[hcho3]

@tdhock Can candidates code in C++? I can add some tests related to C++ core and XGBoost internals

[tdhock]

yes that would be great if you could add tests related to C++ core and xgboost internals

we probably should not accept any student who has not demonstrated C++ coding skills (that is the point of writing these tests)

[hcho3]

@tdhock I'll get to writing tests soon. What would be my responsibilities as a co-mentor? I'm trying to gauge my time commitment. Also, are mentees all hosted on the Google campus?

[tdhock]

no students are hosted on google campus --everything is online

mentor responsibility is basically for one of us to be available for a 1 hour skype call each week during the summer, and to answer student questions via email https://developers.google.com/open-source/gsoc/help/responsibilities

[hcho3]

@tdhock Got it. Thanks for clarification. I'd be more than happy to help.

[hcho3]

@tdhock What's the deadline for the tests? When do you want them by?

[tdhock]

asap! students need to do the tests in the next week or two, then work on an app to submit to google due Apr 7

[hcho3]

Got it. I'll try to post tests as early as possible. Expect them no later than in ~three~ two days from now.

[tdhock]

I added a test about using the functionality in the current xgboost package, but it would be great if you could add some tests about C++ coding / xgboost internals

[thvasilo]

Linking back to a related topic on the forum for left-truncated data: https://discuss.xgboost.ai/t/support-for-left-truncated-data-time-dependent-covariates-for-cox-regression/651

[hcho3]

@tdhock Almost there, it should be done by tomorrow

[hcho3]

@tdhock I've put up five tests (three Easy, one Medium, one Hard). Can you review?

[tdhock]

the first five easy questions are a bit theoretical (gsoc test usually tend to be more practical coding exercises) but for this project I think that is OK, as the student should understand the theory.

median and hard tests look great.

thanks for your help and let's hope we find a good student.

[hcho3]

@tdhock Maybe we should move the binary classification question to "Medium." What do you think?

[tdhock]

either way is fine with me

[hcho3]

Done. Let's hope for the best.

[hcho3]

Closing this in favor of #4491