auto_tune_xgboost {autostats}

auto_tune_xgboost

Description

Automatically tunes an xgboost model using grid or bayesian optimization

Usage

auto_tune_xgboost(
  .data,
  formula,
  tune_method = c("grid", "bayes"),
  event_level = c("first", "second"),
  n_fold = 5L,
  n_iter = 100L,
  seed = 1,
  save_output = FALSE,
  parallel = TRUE,
  trees = tune::tune(),
  min_n = tune::tune(),
  mtry = tune::tune(),
  tree_depth = tune::tune(),
  learn_rate = tune::tune(),
  loss_reduction = tune::tune(),
  sample_size = tune::tune(),
  stop_iter = tune::tune(),
  counts = FALSE,
  tree_method = c("auto", "exact", "approx", "hist", "gpu_hist"),
  monotone_constraints = 0L,
  num_parallel_tree = 1L,
  lambda = 1,
  alpha = 0,
  scale_pos_weight = 1,
  verbosity = 0L
)

Arguments

.data

dataframe

formula

formula

tune_method

method of tuning. defaults to grid

event_level

for binary classification, which factor level is the positive class. specify "second" for second level

n_fold

integer. n folds in resamples

n_iter

n iterations for tuning (bayes); paramter grid size (grid)

seed

seed

save_output

FASLE. If set to TRUE will write the output as an rds file

parallel

default TRUE; If set to TRUE, will enable parallel processing on resamples for grid tuning

trees

# Trees (xgboost: nrounds) (type: integer, default: 500L)

min_n

Minimal Node Size (xgboost: min_child_weight) (type: integer, default: 2L); [typical range: 2-10] Keep small value for highly imbalanced class data where leaf nodes can have smaller size groups. Otherwise increase size to prevent overfitting outliers.

mtry

# Randomly Selected Predictors; defaults to .75; (xgboost: colsample_bynode) (type: numeric, range 0 - 1) (or type: integer if count = TRUE)

tree_depth

Tree Depth (xgboost: max_depth) (type: integer, default: 7L); Typical values: 3-10

learn_rate

Learning Rate (xgboost: eta) (type: double, default: 0.05); Typical values: 0.01-0.3

loss_reduction

Minimum Loss Reduction (xgboost: gamma) (type: double, default: 1.0); range: 0 to Inf; typical value: 0 - 20 assuming low-mid tree depth

sample_size

Proportion Observations Sampled (xgboost: subsample) (type: double, default: .75); Typical values: 0.5 - 1

stop_iter

# Iterations Before Stopping (xgboost: early_stop) (type: integer, default: 15L) only enabled if validation set is provided

counts

if TRUE specify mtry as an integer number of cols. Default FALSE to specify mtry as fraction of cols from 0 to 1

tree_method

xgboost tree_method. default is auto. reference: tree method docs

monotone_constraints

an integer vector with length of the predictor cols, of -1, 1, 0 corresponding to decreasing, increasing, and no constraint respectively for the index of the predictor col. reference: monotonicity docs.

num_parallel_tree

should be set to the size of the forest being trained. default 1L

lambda

[default=.5] L2 regularization term on weights. Increasing this value will make model more conservative.

alpha

[default=.1] L1 regularization term on weights. Increasing this value will make model more conservative.

scale_pos_weight

[default=1] Control the balance of positive and negative weights, useful for unbalanced classes. if set to TRUE, calculates sum(negative instances) / sum(positive instances). If first level is majority class, use values < 1, otherwise normally values >1 are used to balance the class distribution.

verbosity

[default=1] Verbosity of printing messages. Valid values are 0 (silent), 1 (warning), 2 (info), 3 (debug).

Details

Default is to tune all 7 xgboost parameters. Individual parameter values can be optionally fixed to reduce tuning complexity.

Value

workflow object

Examples

iris %>%
 framecleaner::create_dummies() -> iris1

iris1 %>%
 tidy_formula(target = Petal.Length) -> petal_form

iris1 %>%
 rsample::initial_split() -> iris_split

iris_split %>%
 rsample::analysis() -> iris_train

iris_split %>%
 rsample::assessment() -> iris_val

## Not run: 
iris_train %>%
 auto_tune_xgboost(formula = petal_form, n_iter = 10,
 parallel = FALSE, tune_method = "grid", mtry = .5) -> xgb_tuned

xgb_tuned %>%
 parsnip::fit(iris_train) %>%
 parsnip::extract_fit_engine() -> xgb_tuned_fit

xgb_tuned_fit %>%
 tidy_predict(newdata = iris_val, form = petal_form) -> iris_val1

## End(Not run)