Working with lightgbm and catboost

library(tidymodels)
library(treesnip)
data("diamonds", package = "ggplot2")
diamonds <- diamonds %>% sample_n(1000)
# vfold resamples
diamonds_splits <- vfold_cv(diamonds, v = 5)

Model spec

Both lightgbm and catboost are engines for parsnip::boost_tree(). All main hyperparameters are available, except loss_reduction for catboost.

model_spec <- boost_tree(mtry = 5, trees = 500) %>% set_mode("regression")

# model specs
lightgbm_model <- model_spec %>% set_engine("lightgbm", nthread = 6)
catboost_model <- model_spec %>% set_engine("catboost", nthread = 6)

#workflows
lightgbm_wf <- workflow() %>% add_model(lightgbm_model)
catboost_wf <- workflow() %>% add_model(catboost_model)

Recipes and categorical features

Differently from xgboost, lightgbm and catboost deals with nominal columns natively. No step_dummy() or any other encoding required.

Encodings benchmark

In fact, avoiding step_dummy() seems to be a good idea when using lightgbm or catboost.

rec_ordered <- recipe(price ~ ., data = diamonds) 
rec_dummy <- rec_ordered %>% step_dummy(all_nominal())
rec_factor <- rec_ordered %>% step_unorder(all_nominal())

# glimpse(juice(prep(rec_factor)))

lightgbm_fit_ordered <- fit_resamples(add_recipe(lightgbm_wf, rec_ordered), resamples = diamonds_splits)
lightgbm_fit_dummy <- fit_resamples(add_recipe(lightgbm_wf, rec_dummy), resamples = diamonds_splits)
lightgbm_fit_factor <- fit_resamples(add_recipe(lightgbm_wf, rec_factor), resamples = diamonds_splits)
catboost_fit_ordered <- fit_resamples(add_recipe(catboost_wf, rec_ordered), resamples = diamonds_splits)
catboost_fit_dummy <- fit_resamples(add_recipe(catboost_wf, rec_dummy), resamples = diamonds_splits)
catboost_fit_factor <- fit_resamples(add_recipe(catboost_wf, rec_factor), resamples = diamonds_splits)

bind_rows(
  collect_metrics(lightgbm_fit_ordered) %>% mutate(model = "lightgbm", encode = "ordered"),
  collect_metrics(lightgbm_fit_dummy) %>% mutate(model = "lightgbm", encode = "dummy"),
  collect_metrics(lightgbm_fit_factor) %>% mutate(model = "lightgbm", encode = "factor"),
  collect_metrics(catboost_fit_ordered) %>% mutate(model = "catboost", encode = "ordered"),
  collect_metrics(catboost_fit_dummy) %>% mutate(model = "catboost", encode = "dummy"),
  collect_metrics(catboost_fit_factor) %>% mutate(model = "catboost", encode = "factor")
) %>%
  filter(.metric == "rmse") %>%
  ggplot(aes(x = encode, y = mean, group = model, colour = model)) +
  geom_point() +
  geom_errorbar(aes(ymin = mean - std_err, ymax = mean + std_err), width = 0.2) +
  geom_line() +
  labs(y = "RMSE (the lower the better)") +
  theme_minimal(20)