LightRFHyperparameters

Properties

boosting

string

Boosting type. 'rf' is what makes LightGBM a random forest.

learning_rate

number

Learning rate. No effect in 'rf' mode; set for clarity.

bagging_freq

number

Bagging frequency.

early_stopping_rounds

number

Early stopping rounds. -1 disables early stopping.

nrounds

integer | object≥ 1

Number of boosting rounds (trees).

num_leaves

integer | object≥ 1

Maximum number of leaves per tree.

max_depth

integer | object

Maximum tree depth. -1 = no limit.

feature_fraction

number | object | null≤ 1> 0

Fraction of features sampled per tree. NULL = sqrt(n_features)/n_features for classification, 0.33 for regression.

bagging_fraction

number | object≤ 1> 0

Fraction of cases sampled per tree (bagging fraction).

lambda_l1

number | object≥ 0

L1 regularization.

lambda_l2

number | object≥ 0

L2 regularization.

max_cat_threshold

integer | object≥ 1

Maximum number of split points for categorical features.

min_data_per_group

integer | object≥ 1

Minimum number of cases per categorical group.

linear_tree

boolean | object

Fit linear models at leaves.

ifw

boolean | object

Inverse frequency weighting of outcome classes.

objective

string | null

LightGBM objective. NULL = set from outcome type.

device_type

string

Compute device.

one of"cpu""gpu""cuda"

tree_learner

string

Tree learner type.

one of"serial""feature""data""voting"

force_col_wise

boolean

Force column-wise histogram building (CPU only).

alpha

number | object | null> 0

Huber delta or quantile level. NULL = LightGBM's default of 0.9. Applies only when objective is huber or quantile.

tweedie_variance_power

number | object | null≥ 1< 2

Tweedie power: 1 is Poisson-like, near 2 is gamma-like. NULL = LightGBM's default of 1.5. Applies only when objective is tweedie.

fair_c

number | object | null> 0

Fair loss scale. NULL = LightGBM's default of 1. Applies only when objective is fair.

poisson_max_delta_step

number | object | null> 0

Step cap safeguarding Poisson optimization. NULL = LightGBM's default of 0.7. Applies only when objective is poisson.

sigmoid

number | object | null> 0

Sigmoid slope, for a binary or one-vs-all objective. NULL = LightGBM's default of 1.

boost_from_average

boolean | null

Start from the outcome's mean, for a regression, binary or cross-entropy objective. NULL = LightGBM's default of TRUE.

reg_sqrt

boolean | null

Fit the square root of the outcome and square the prediction back, for a regression objective. NULL = LightGBM's default of FALSE.

min_data_in_leaf

integer | object≥ 0

Minimum number of cases in a leaf.

min_sum_hessian_in_leaf

number | object≥ 0

Minimum sum of hessians in a leaf: the smooth counterpart of a case count.

min_gain_to_split

number | object≥ 0

Minimum loss reduction a split must buy to be made.

max_delta_step

number | object

Cap on each leaf's output; 0 or less leaves it uncapped.

path_smooth

number | object≥ 0

Smooth each leaf towards its parent, more strongly the fewer cases it holds.

extra_trees

boolean | object

Choose one split threshold per feature at random rather than the best.

extra_seed

integer

Random seed for the thresholds `extra_trees` draws.

linear_lambda

number | object | null≥ 0

L2 penalty on the linear models fitted at the leaves. NULL = LightGBM's default of 0. Applies only when linear_tree is TRUE.

max_bin

integer | object≥ 2

Most bins a feature is discretized into. Fewer is faster and more regularized.

min_data_in_bin

integer | object≥ 1

Minimum number of cases per bin.

use_missing

boolean

Handle missing values. FALSE ignores them.

zero_as_missing

boolean

Treat zeros as missing, rather than only NA.

cat_l2

number | object≥ 0

L2 regularization on a categorical split.

cat_smooth

number | object≥ 0

Smoothing over categorical levels, which reduces the effect of rare ones.

max_cat_to_onehot

integer | object≥ 1

Below this many levels, split one-versus-rest rather than by partition.

deterministic

boolean

Force reproducible results at some cost in speed. Requires a forced histogram direction.

objective_seed

integer

Random seed for the objective, where it draws.

force_row_wise

boolean

Force row-wise histogram building. Cheaper in memory than column-wise, slower per iteration.

histogram_pool_size

number

Memory cap in MB for the histogram cache; -1 leaves it uncapped.

top_k

integer | object | null≥ 1

Features each machine votes for in the voting tree learner. NULL = LightGBM's default of 20. Applies only when tree_learner is voting.

pos_bagging_fraction

number | object≤ 1> 0

Fraction of positive cases sampled per tree, for a binary outcome.

neg_bagging_fraction

number | object≤ 1> 0

Fraction of negative cases sampled per tree, for a binary outcome.

bagging_seed

integer

Random seed for bagging.

feature_fraction_bynode

number | object≤ 1> 0

Fraction of features sampled at each node, on top of the per-tree fraction.

feature_fraction_seed

integer

Random seed for feature sampling.

cegb_tradeoff

number | object≥ 0

Overall multiplier on every cost-efficiency penalty below.

cegb_penalty_split

number | object≥ 0

Cost charged per split, in units of loss reduction.

cegb_penalty_feature_lazy

number[] | number[][] | null≥ 1 items≥ 1 items

Per-feature cost charged the first time a case uses it, one value per feature.

cegb_penalty_feature_coupled

number[] | number[][] | null≥ 1 items≥ 1 items

Per-feature cost charged the first time any case uses it, one value per feature.

use_quantized_grad

boolean

Discretize gradients before finding splits. Faster, and changes the fit.

num_grad_quant_bins

integer | object | null≥ 2

Bins the gradients are discretized into. NULL = LightGBM's default of 4. Applies only when use_quantized_grad is TRUE.

quant_train_renew_leaf

boolean | null

Recompute leaf values from the exact gradients after each iteration. NULL = LightGBM's default of FALSE. Applies only when use_quantized_grad is TRUE.

stochastic_rounding

boolean | null

Round gradients stochastically rather than to nearest. NULL = LightGBM's default of TRUE. Applies only when use_quantized_grad is TRUE.

monotone_constraints

array | null≥ 1 items

Per-feature monotonicity: 1 increasing, -1 decreasing, 0 unconstrained. Must have one value per feature.

monotone_constraints_method

string | null

How monotonicity is enforced: 'basic' is fastest, 'advanced' least restrictive. Applies only when monotone_constraints is -1 or 1.

one of"basic""intermediate""advanced"null

monotone_penalty

number | object | null≥ 0

Depth penalty on splits that would break monotonicity. NULL = LightGBM's default of 0. Applies only when monotone_constraints is -1 or 1.

interaction_constraints

string | null

Feature groups permitted to interact, in LightGBM's own bracketed form, e.g. "[0,1],[2,3]".

feature_contri

array | null≥ 1 items

Per-feature multiplier on split gain, one value per feature. Must have one value per feature.

early_stopping_min_delta

number | object≥ 0

Smallest improvement that resets the early-stopping counter.

Relationships