Boosted Trees, the objective parameter and count:poisson

User 2568 | 6/20/2016, 10:04:59 PM

I noticed that xboost, which I believe is the Graph Lab impelmenation of boosted trees, takes a range objective function. It seems that some of these like reg:linear, reg:logistic would are set depending on whether I use a boosted tree classifier or a regression . However, there are some other like count:poisson, multi:softmax, multi:softprob, custom rank:pairwise or potentially a customised loss function that are not exposed by Graph Lab create.

Any plans to expose the objective function in later releases? I


User 940 | 6/22/2016, 8:08:20 PM

Hi @Kevin_McIsaac ,

I just checked with the team, and it seems like it is on our roadmap but not in the immediate future.

I hope this helps!

Cheers! -Piotr