gbm package and quantile regression

Question

Could someone please point out the correct use of the quantile distribution option in the gbm package? This:

library(datasets)
library(gbm)
library(caret)

set.seed(42)
rm(list = ls())

model <- gbm(Petal.Width ~ Petal.Length

                        , distribution = list(name = "quantile", alpha = 0.4)
                        , data = iris
                        , n.trees = number_of_trees
                        , interaction.depth = 3
                        , shrinkage = 0.01,
                        , n.minobsinnode = 10
    )
model

Does not work. I get:

Error in if (!is.element(distribution$name, getAvailableDistributions())) { : 
  argument is of length zero
Error: object 'model' not found

Thanks!

Answer 1

This was a bug in gbm, as reported in these GitHub issues: #29, #27. It was fixed in this commit. Until they get the new version on CRAN, you can do quantile regression with the GitHub development version:

devtools::install_github("gbm-developers/gbm")
#> Downloading GitHub repo gbm-developers/gbm@master
#> from URL https://api.github.com/repos/gbm-developers/gbm/zipball/master
#> Installing gbm
#> '/usr/lib/R/bin/R' --no-site-file --no-environ --no-save --no-restore  \
#>   --quiet CMD INSTALL  \
#>   '/tmp/Rtmp4acgli/devtools55756447fca5/gbm-developers-gbm-0e07a6b'  \
#>   --library='/home/duckmayr/R/x86_64-pc-linux-gnu-library/3.5' --install-tests
#> 
#> Reloading installed gbm
#> Loaded gbm 2.1.4.9000
library(datasets)
library(gbm)
# library(caret) # this package isn't used

set.seed(42)
rm(list = ls())

model <- gbm(Petal.Width ~ Petal.Length

             , distribution = list(name = "quantile", alpha = 0.4)
             , data = iris
             , n.trees = 3 # number_of_trees -- this variable isn't given by OP
             , interaction.depth = 3
             , shrinkage = 0.01,
             , n.minobsinnode = 10
)
model
#> gbm(formula = Petal.Width ~ Petal.Length, distribution = list(name = "quantile", 
#>     alpha = 0.4), data = iris, n.trees = 3, interaction.depth = 3, 
#>     n.minobsinnode = 10, shrinkage = 0.01)
#> A gradient boosted model with quantile loss function.
#> 3 iterations were performed.
#> There were 1 predictors of which 1 had non-zero influence.

but not the CRAN version:

install.packages("gbm")
#> Installing package into '/home/duckmayr/R/x86_64-pc-linux-gnu-library/3.5'
#> (as 'lib' is unspecified)
library(datasets)
library(gbm)
#> Loaded gbm 2.1.4
# library(caret) # this package isn't used

set.seed(42)
rm(list = ls())

model <- gbm(Petal.Width ~ Petal.Length

             , distribution = list(name = "quantile", alpha = 0.4)
             , data = iris
             , n.trees = 3 # number_of_trees -- this variable isn't given by OP
             , interaction.depth = 3
             , shrinkage = 0.01,
             , n.minobsinnode = 10
)
#> Error in if (!is.element(distribution$name, getAvailableDistributions())) {: argument is of length zero
model
#> Error in eval(expr, envir, enclos): object 'model' not found

The issue was caused by this bit of code:

distribution <- if (missing(distribution)) {
  if (missing(distribution)) {
    y <- data[, all.vars(formula)[1L], drop = TRUE]
    guessDist(y) 
  } else if (is.character(distribution)) { 
    distribution <- list(name = distribution) 
  } 
}

You'll notice they forgot at some point to handle the case where users pass a named list like the documentation said they could. But, now that bit of code is fixed:

if (missing(distribution)) {
  y <- data[, all.vars(formula)[1L], drop = TRUE]
  distribution <- guessDist(y) 
}

if (is.character(distribution)) { 
  distribution <- list(name = distribution) 
}

That way if distribution is already a list, it is left undisturbed now.