feat: add quantile regression to ranger

DESCRIPTION

@@ -47,6 +47,8 @@ Suggests:
     rmarkdown,
     testthat (>= 3.0.0),
     xgboost (>= 1.6.0)
+Remotes:
+    mlr-org/mlr3@predict_type_quantile
 Config/testthat/edition: 3
 Encoding: UTF-8
 NeedsCompilation: no

R/LearnerRegrRanger.R

@@ -43,7 +43,6 @@ LearnerRegrRanger = R6Class("LearnerRegrRanger",
         num.threads                  = p_int(1L, default = 1L, tags = c("train", "predict", "threads")),
         num.trees                    = p_int(1L, default = 500L, tags = c("train", "predict", "hotstart")),
         oob.error                    = p_lgl(default = TRUE, tags = "train"),
-        quantreg                     = p_lgl(default = FALSE, tags = "train"),
         regularization.factor        = p_uty(default = 1, tags = "train"),
         regularization.usedepth      = p_lgl(default = FALSE, tags = "train"),
         replace                      = p_lgl(default = TRUE, tags = "train"),
@@ -64,7 +63,7 @@ LearnerRegrRanger = R6Class("LearnerRegrRanger",
       super$initialize(
         id = "regr.ranger",
         param_set = ps,
-        predict_types = c("response", "se"),
+        predict_types = c("response", "se", "quantile"),
         feature_types = c("logical", "integer", "numeric", "character", "factor", "ordered"),
         properties = c("weights", "importance", "oob_error", "hotstart_backward"),
         packages = c("mlr3learners", "ranger"),
@@ -111,6 +110,10 @@ LearnerRegrRanger = R6Class("LearnerRegrRanger",
         pv$keep.inbag = TRUE # nolint
       }
 
+      if (self$predict_type == "quantile") {
+        pv$quantreg = TRUE # nolint
+      }
+
       invoke(ranger::ranger,
         dependent.variable.name = task$target_names,
         data = task$data(),
@@ -123,8 +126,23 @@ LearnerRegrRanger = R6Class("LearnerRegrRanger",
       pv = self$param_set$get_values(tags = "predict")
       newdata = ordered_features(task, self)
 
-      prediction = invoke(predict, self$model, data = newdata, type = self$predict_type, .args = pv)
-      list(response = prediction$predictions, se = prediction$se)
+      type = switch(
+        self$predict_type,
+        response = "response",
+        se = "se",
+        quantile = "quantiles"
+      )
+
+      prediction = invoke(predict, self$model, data = newdata, type = type, quantiles = private$.quantile, .args = pv)
+
+      if (type == "quantiles") {
+        response = prediction$predictions[, which(private$.quantile == private$.quantile_response)]
+        quantile = prediction$predictions
+        attr(quantile, "probs") = private$.quantile
+        list(response = response, quantile = quantile)
+      } else {
+        list(response = prediction$predictions, se = prediction$se)
+      }
     },
 
     .hotstart = function(task) {

inst/paramtest/test_paramtest_regr.ranger.R

@@ -17,7 +17,8 @@ test_that("regr.ranger", {
     "status.variable.name", # handled via mlr3
     "classification", # handled via mlr3
     "mtry.ratio", # custom hyperpar
-    "time.interest" # survival only
+    "time.interest", # survival only
+    "quantreg" # handled by predict_type
   )
 
   ParamTest = run_paramtest(learner, fun, exclude, tag = "train")

tests/testthat/test_regr_ranger.R

@@ -66,3 +66,17 @@ test_that("default_values", {
   values = default_values(learner, search_space, task)
   expect_names(names(values), permutation.of =  c("replace", "sample.fraction", "num.trees", "mtry.ratio"))
 })
+
+test_that("quantile prediction", {
+  learner = mlr3::lrn("regr.ranger", num.trees = 100, predict_type = "quantile")
+  learner$quantiles = c(0.1, 0.5, 0.9)
+  learner$quantile_response = 0.5
+  task = tsk("mtcars")
+
+  learner$train(task)
+  pred = learner$predict(task)
+  expect_matrix(pred$quantile)
+  tab = as.data.table(pred)
+  expect_names(names(tab), must.include = c("q0.1", "q0.5", "q0.9", "response"))
+  expect_equal(tab$response, tab$q0.5)
+})