search

harrysouthworth / gbm

Gradient boosted models
Other
106 stars 27 forks source link

weighted quantile regression #10

Closed searchivarius closed 9 years ago

searchivarius commented 10 years ago

Dear Harry,

This is to enable a weighted quantile regression. We used this code in one of our projects (on top of gbm 2.1).

We also created an example to ensure that the output of the truly weighted regression roughly corresponds to the output of unweighted quantile regression where observations are repeated according to the weights.

This example can be downloaded from http://boytsov.info/tmp/examplefilesforgbmquantile.tar.gz

An important note: the R-wrapper in the master branch appears to be broken. We can run this example in gbm 2.1-0.3 (with a minimum change of C++ code, namely, using unsigned long * aiNodeAssign instead of the vector std::vector& aiNodeAssign). Yet, the example crashes when we run using the master branch.

Furthermore, the unweighted quantile regression crashes even if we use the unmodified master branch.

Leo & Anna

harrysouthworth commented 10 years ago

Thanks for this.

Because of the issue you mention, and because of a slew of other issues that has arrived in my inbox, I'm going to hold off merging this one until I get a decent amount of time to look at what's going on.

Thanks again. It's appreciated. Harry

searchivarius commented 10 years ago

No problem, thanks!

Neil-Schneider commented 9 years ago

With the acceptance of #37 into the master code base, i would recommend closing this pull request to avoid future confusion.

harrysouthworth commented 9 years ago

Ok, thanks. Harry