`mtry_prop` vs `mtry` in boosted trees

[frankiethull]

hi all -

I was curious why the catboost tuning parameter is called mtry_prop instead of mtry?

For lightgbm I have been doing, counts = FALSE, and set mtry is a proportion when cross-validating. I actually found that method jives better with future and future.mirai when creating a parallel plan.

... # setting it here, then tuning mtry is a proportion:
parsnip::set_engine("lightgbm", counts = FALSE)

I prefer the mtry_prop over the mtry count method. But I am just curious if this parameter will stay mtry for xgboost + lightgbm but mtry_prop for catboost. This will change how we build out our catboosters 🐱, appreciate the help!

[frankiethull]

Ah, so I reran with mtry_prop and see a different error. I guess I am unsure what the tunable params are for catboost! Let me know if there's documentation on these, I checked here but catboost is not listed (yet).

thanks again!

################################################################
#                       cat cross-val
################################################################
library(catboost)
library(parsnip)
library(bonsai)
library(rsample)
library(recipes)
#> Loading required package: dplyr
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
#> 
#> Attaching package: 'recipes'
#> The following object is masked from 'package:stats':
#> 
#>     step
library(stacks)
#> Registered S3 method overwritten by 'butcher':
#>   method                 from    
#>   as.character.dev_topic generics
library(yardstick)
library(workflows)
library(tune)
# taken from stacks --------------------------------
# https://stacks.tidymodels.org/articles/basics.html
data("tree_frogs")

# subset the data
tree_frogs <- tree_frogs |>
  filter(!is.na(latency)) |>
  select(-c(clutch, hatched))
# some setup: resampling and a basic recipe
set.seed(1)
tree_frogs_split <- initial_split(tree_frogs)
tree_frogs_train <- training(tree_frogs_split)
tree_frogs_test  <- testing(tree_frogs_split)

set.seed(1)
folds <- rsample::vfold_cv(tree_frogs_train, v = 5)

tree_frogs_rec <-
  recipe(latency ~ ., data = tree_frogs_train)

metric <- metric_set(rmse)

ctrl_grid <- control_stack_grid()
ctrl_res <- control_stack_resamples()


### ~~
### change the knn to catboost, test mtry tuning:
cat_spec <-
  boost_tree(
    mode = "regression",
    mtry = tune("mtry")
  ) |>
  set_engine("catboost")

# extend the recipe
#
tree_frogs_rec <-
 recipe(latency ~ ., data = tree_frogs_train)

cat_rec <-
  tree_frogs_rec |>
  step_dummy(all_nominal_predictors()) |>
  step_zv(all_predictors()) |>
  step_impute_mean(all_numeric_predictors()) |>
  step_normalize(all_numeric_predictors())

#cat_rec

cat_wflow <-
  workflow() |>
  add_model(cat_spec) |>
  add_recipe(cat_rec)


# tune mtry and fit to the 5-fold cv
set.seed(2020)
cat_res <-
  tune_grid(
    cat_wflow,
    resamples = folds,
    metrics = metric,
    grid = 4,
    control = ctrl_grid
  )
#> Error in `tune_grid()`:
#> ! The parameter `mtry` was marked with `tune()`, though will not be
#>   tuned.
#> ℹ This usually means that the current modeling engine `catboost` does not
#>   support tuning `mtry`.

# saw this and dials::mtry_prop so thought it was mapped differently. 
# https://github.com/tidymodels/bonsai/blob/f63ff2eaa8a225540a0559291bc0583ee08a83c8/R/catboost_data.R#L187
cat_spec <-
  boost_tree(
    mode = "regression",
    mtry_prop = tune("mtry_prop")
  ) |>
  set_engine("catboost")
#> Error in boost_tree(mode = "regression", mtry_prop = tune("mtry_prop")): unused argument (mtry_prop = tune("mtry_prop"))

^{Created on 2025-08-01 with reprex v2.1.1}