LightGBM gradient-boosting backend for xplus

Gradient-boosted trees via lightgbm::lgb.train(). Binary targets use the "binary" objective, fractional pseudo-labels "regression" (clipped to [0, 1] on prediction). Hyperparameters use parsnip-harmonized names. fit_predict returns out-of-fold predictions via lgb.cv on the folds xplus() assigns, so the classification cutoff is calibrated on honest scores rather than memorized ones.

Usage

xplus_backend_lightgbm(
  trees = 100,
  learn_rate = 0.1,
  tree_depth = -1,
  min_n = 20,
  mtry = NULL,
  sample_size = 1,
  loss_reduction = 0,
  num.threads = 1,
  seed = NULL
)

Arguments

trees

Number of boosting rounds (lightgbm num_iterations).

learn_rate

Shrinkage step (lightgbm learning_rate).

tree_depth

Maximum tree depth (lightgbm max_depth; -1 unbounded).

min_n

Minimum observations per leaf (lightgbm min_data_in_leaf).

mtry

Number of features sampled per iteration; NULL uses all. Converted to feature_fraction as a proportion of ncol(x).

sample_size

Row subsampling fraction per iteration (lightgbm bagging_fraction; enables bagging when below 1).

loss_reduction

Minimum gain to split (lightgbm min_gain_to_split).

num.threads

OpenMP threads (lightgbm num_threads).

seed

Optional seed passed to each fit; NULL draws from R's RNG per fit so that xplus(seed = ...) remains reproducible.

Value

An "xplus_backend" object.

See also

xplus_backend(), xplus(), xplus_backend_xgboost()

Examples

if (FALSE) { # \dontrun{
fit <- xplus(x, y, backend = xplus_backend_lightgbm(trees = 200))
} # }