boosters.GBDTConfig#

class boosters.GBDTConfig#

Bases: object

Main configuration for GBDT model.

This is the primary configuration class for gradient boosted decision trees. All parameters are flat (no nested config objects) matching the core Rust API.

# Arguments

n_estimators - Number of boosting rounds (trees to train). Default: 100.
learning_rate - Step size shrinkage (0.01 - 0.3 typical). Default: 0.3.
objective - Loss function for training. Default: Objective.Squared().
metric - Evaluation metric. None uses objective’s default.
growth_strategy - Tree growth strategy. Default: GrowthStrategy.Depthwise.
max_depth - Maximum tree depth (only for depthwise). Default: 6.
n_leaves - Maximum leaves (only for leafwise). Default: 31.
max_onehot_cats - Max categories for one-hot encoding. Default: 4.
l1 - L1 regularization on leaf weights. Default: 0.0.
l2 - L2 regularization on leaf weights. Default: 1.0.
min_gain_to_split - Minimum gain required to make a split. Default: 0.0.
min_child_weight - Minimum sum of hessians in a leaf. Default: 1.0.
min_samples_leaf - Minimum samples in a leaf. Default: 1.
subsample - Row subsampling ratio per tree. Default: 1.0.
colsample_bytree - Column subsampling per tree. Default: 1.0.
colsample_bylevel - Column subsampling per level. Default: 1.0.
linear_leaves - Enable linear models in leaves (experimental). Default: False.
linear_l2 - L2 regularization for linear coefficients. Default: 0.01.
linear_l1 - L1 regularization for linear coefficients. Default: 0.0.
early_stopping_rounds - Stop if no improvement for this many rounds.
seed - Random seed for reproducibility. Default: 42.

# Example (Python)

```text config = GBDTConfig(

n_estimators=500, learning_rate=0.1, objective=Objective.logistic(), max_depth=6, l2=1.0,

)#

classmethod __new__(*args, **kwargs)#

binning_sample_cnt#: Number of samples for computing bin boundaries (for large datasets).

cache_size#

Type:: Histogram cache size (number of slots). Default

colsample_bylevel#: Column subsampling ratio per level.

colsample_bytree#: Column subsampling ratio per tree.

early_stopping_rounds#: Early stopping rounds (None = disabled).

enable_bundling#

true. Bundles sparse/one-hot features to reduce memory and speed up training.

Type:: Enable exclusive feature bundling for sparse data. Default

growth_strategy#: Growth strategy for tree building.

l1#: L1 regularization on leaf weights.

l2#: L2 regularization on leaf weights.

learning_rate#: Learning rate (step size shrinkage).

linear_coefficient_threshold#: Threshold for pruning small coefficients.

linear_l1#: L1 regularization for linear coefficients.

linear_l2#: L2 regularization for linear coefficients.

linear_leaves#: Enable linear models in leaves.

linear_max_features#: Maximum features in linear model per leaf.

linear_max_iterations#: Maximum coordinate descent iterations for linear leaves.

linear_min_samples#: Minimum samples required to fit linear model in leaf.

linear_skip_first_n_trees#: Number of initial trees to skip linear leaf fitting. Default is 1 (first tree has homogeneous gradients). Set to 0 to enable from first tree.

linear_tolerance#: Convergence tolerance for linear leaves.

linear_use_global_features#: Use global features instead of path features for linear models. When true, uses the top-k most frequently split features for all leaves. This can improve extrapolation by ensuring important features are always included.

max_bins#: Maximum bins per feature for binning (1-256).

max_categorical_cardinality#: Max cardinality to auto-detect as categorical. Features with ≤ this many unique integer values may be treated as categorical.

max_depth#: Maximum depth of tree (for depthwise growth).

max_onehot_cats#: Maximum categories for one-hot encoding categorical splits.

metric#: Get the evaluation metric (or None).

min_child_weight#: Minimum sum of hessians required in a leaf.

min_gain_to_split#: Minimum gain required to make a split.

min_samples_leaf#: Minimum number of samples required in a leaf.

n_estimators#: Number of boosting rounds.

n_leaves#: Maximum number of leaves (for leafwise growth).

objective#: Get the objective function.

seed#: Random seed.

sparsity_threshold#: Sparsity threshold (fraction of zeros to use sparse storage). Features with density ≤ (1 - threshold) are considered sparse.

subsample#: Row subsampling ratio per tree.

verbosity#: Verbosity level for training output.