search

Evovest / EvoTrees.jl

Boosted trees in Julia
https://evovest.github.io/EvoTrees.jl/dev/
Apache License 2.0
174 stars 20 forks source link

Feature Request MultiQuantile #269

Open 1991jhf opened 2 weeks ago

1991jhf commented 2 weeks ago

It would be nice to have multi-quantile regression for approximating histogram in one go. similar to https://catboost.ai/en/docs/concepts/loss-functions-regression#MultiQuantile

jeremiedb commented 2 weeks ago

Thanks for the pointer.

The part I found tricky in supporting quantiles is that as the loss isn't differentiable per se, I ended up tracking full vector in observations/targets in order to compute the leaf prediction values. Having a reliable gradient based proxy would be preferable. Notably for maintenance as I for now don't have much use case with such loss objectives. If you can point to a gradient based approach for estimating the loss/gain and associated leaf predictions, it should be fairly strightforward to add support for this loss.