enabled quantile regression in gbm

NEWS.md

@@ -11,6 +11,7 @@
   See `?regr.randomForest` for more details.  
   `regr.ranger` relies on the functions provided by the package ("jackknife" and "infjackknife" (default))  
   (@jakob-r, #1784)
+- `regr.gbm` now supports `quantile distribution` (@bthieurmel, #2603)
 
 ## functions - general
 - `getClassWeightParam()` now also works for Wrapper* Models and ensemble models (@ja-thomas, #891)

R/RLearner_regr_gbm.R

@@ -4,7 +4,7 @@ makeRLearner.regr.gbm = function() {
     cl = "regr.gbm",
     package = "gbm",
     par.set = makeParamSet(
-      makeDiscreteLearnerParam(id = "distribution", default = "gaussian", values = c("gaussian", "laplace", "poisson", "tdist")),
+      makeDiscreteLearnerParam(id = "distribution", default = "gaussian", values = c("gaussian", "laplace", "poisson", "tdist", "quantile")),
       # FIXME default for distribution in gbm() is bernoulli
       makeIntegerLearnerParam(id = "n.trees", default = 100L, lower = 1L),
       makeIntegerLearnerParam(id = "cv.folds", default = 0L),
@@ -13,6 +13,8 @@ makeRLearner.regr.gbm = function() {
       makeNumericLearnerParam(id = "shrinkage", default = 0.001, lower = 0),
       makeNumericLearnerParam(id = "bag.fraction", default = 0.5, lower = 0, upper = 1),
       makeNumericLearnerParam(id = "train.fraction", default = 1, lower = 0, upper = 1),
+      makeNumericLearnerParam(id = "alpha", default = 0.5, lower = 0, upper = 1,
+                              requires = quote(distribution == "quantile")),
       makeLogicalLearnerParam(id = "keep.data", default = TRUE, tunable = FALSE),
       makeLogicalLearnerParam(id = "verbose", default = FALSE, tunable = FALSE)
     ),
@@ -28,13 +30,25 @@ makeRLearner.regr.gbm = function() {
 #' @export
 trainLearner.regr.gbm = function(.learner, .task, .subset, .weights = NULL, ...) {
   f = getTaskFormula(.task)
-  if (is.null(.weights)) {
-    f = getTaskFormula(.task)
-    gbm::gbm(f, data = getTaskData(.task, .subset), ...)
+
+  params = list(...)
+  if("alpha" %in% names(params)) {
+    alpha = params$alpha
+    params$alpha = NULL
   } else {
-    f = getTaskFormula(.task)
-    gbm::gbm(f, data = getTaskData(.task, .subset), weights = .weights, ...)
+    alpha = 0.5
+  }
+  if(params$distribution %in% "quantile") {
+    params$distribution = list(name = "quantile", alpha = alpha)
   }
+  params$formula = f
+  params$data = getTaskData(.task, .subset)
+
+  if (!is.null(.weights)) {
+    params$weights = .weights
+  }
+
+  do.call(gbm::gbm, params)
 }
 
 #' @export

tests/testthat/test_regr_gbm.R

@@ -5,16 +5,17 @@ test_that("regr_gbm", {
 
   parset.list = list(
     list(),
-    list(n.trees = 600),
-    list(interaction.depth = 2)
+    list(n.trees = 600, distribution = "gaussian"),
+    list(interaction.depth = 2, distribution = "gaussian"),
+    list(distribution = list(name = "quantile", alpha = 0.2))
   )
 
 
   old.predicts.list = list()
 
   for (i in seq_along(parset.list)) {
     parset = parset.list[[i]]
-    pars = list(regr.formula, data = regr.train, distribution = "gaussian")
+    pars = list(regr.formula, data = regr.train)
     pars = c(pars, parset)
     set.seed(getOption("mlr.debug.seed"))
     capture.output({
@@ -25,6 +26,10 @@ test_that("regr_gbm", {
     old.predicts.list[[i]] = p
   }
 
+  #  Different way to pass quantile distribution in mlr
+  parset.list[[4]]$distribution = "quantile"
+  parset.list[[4]]$alpha = 0.2
+
   testSimpleParsets("regr.gbm", regr.df, regr.target, regr.train.inds, old.predicts.list, parset.list)
 })