Skip to contents

Classification LightGBM Learner

Source: R/learner_lightgbm_classif_lightgbm.R
mlr_learners_classif.lightgbm.Rd

Gradient boosting algorithm. Calls lightgbm::lightgbm() from lightgbm. The list of parameters can be found here and in the documentation of lightgbm::lgb.train().

Dictionary

This Learner can be instantiated via lrn():

lrn("classif.lightgbm")

Meta Information

  • Task type: “classif”

  • Predict Types: “response”, “prob”

  • Feature Types: “logical”, “integer”, “numeric”, “factor”

  • Required Packages: mlr3, mlr3extralearners, lightgbm

Parameters

IdTypeDefaultLevelsRange
objectivecharacter-binary, multiclass, multiclassova-
evaluntyped--
verboseinteger1\((-\infty, \infty)\)
recordlogicalTRUETRUE, FALSE-
eval_freqinteger1\([1, \infty)\)
callbacksuntyped--
reset_datalogicalFALSETRUE, FALSE-
boostingcharactergbdtgbdt, rf, dart, goss-
linear_treelogicalFALSETRUE, FALSE-
learning_ratenumeric0.1\([0, \infty)\)
num_leavesinteger31\([1, 131072]\)
tree_learnercharacterserialserial, feature, data, voting-
num_threadsinteger0\([0, \infty)\)
device_typecharactercpucpu, gpu-
seedinteger-\((-\infty, \infty)\)
deterministiclogicalFALSETRUE, FALSE-
data_sample_strategycharacterbaggingbagging, goss-
force_col_wiselogicalFALSETRUE, FALSE-
force_row_wiselogicalFALSETRUE, FALSE-
histogram_pool_sizenumeric-1\((-\infty, \infty)\)
max_depthinteger-1\((-\infty, \infty)\)
min_data_in_leafinteger20\([0, \infty)\)
min_sum_hessian_in_leafnumeric0.001\([0, \infty)\)
bagging_fractionnumeric1\([0, 1]\)
pos_bagging_fractionnumeric1\([0, 1]\)
neg_bagging_fractionnumeric1\([0, 1]\)
bagging_freqinteger0\([0, \infty)\)
bagging_seedinteger3\((-\infty, \infty)\)
bagging_by_querylogicalFALSETRUE, FALSE-
feature_fractionnumeric1\([0, 1]\)
feature_fraction_bynodenumeric1\([0, 1]\)
feature_fraction_seedinteger2\((-\infty, \infty)\)
extra_treeslogicalFALSETRUE, FALSE-
extra_seedinteger6\((-\infty, \infty)\)
max_delta_stepnumeric0\((-\infty, \infty)\)
lambda_l1numeric0\([0, \infty)\)
lambda_l2numeric0\([0, \infty)\)
linear_lambdanumeric0\([0, \infty)\)
min_gain_to_splitnumeric0\([0, \infty)\)
drop_ratenumeric0.1\([0, 1]\)
max_dropinteger50\((-\infty, \infty)\)
skip_dropnumeric0.5\([0, 1]\)
xgboost_dart_modelogicalFALSETRUE, FALSE-
uniform_droplogicalFALSETRUE, FALSE-
drop_seedinteger4\((-\infty, \infty)\)
top_ratenumeric0.2\([0, 1]\)
other_ratenumeric0.1\([0, 1]\)
min_data_per_groupinteger100\([1, \infty)\)
max_cat_thresholdinteger32\([1, \infty)\)
cat_l2numeric10\([0, \infty)\)
cat_smoothnumeric10\([0, \infty)\)
max_cat_to_onehotinteger4\([1, \infty)\)
top_kinteger20\([1, \infty)\)
monotone_constraintsuntypedNULL-
monotone_constraints_methodcharacterbasicbasic, intermediate, advanced-
monotone_penaltynumeric0\([0, \infty)\)
feature_contriuntypedNULL-
forcedsplits_filenameuntyped""-
refit_decay_ratenumeric0.9\([0, 1]\)
cegb_tradeoffnumeric1\([0, \infty)\)
cegb_penalty_splitnumeric0\([0, \infty)\)
cegb_penalty_feature_lazyuntyped--
cegb_penalty_feature_coupleduntyped--
path_smoothnumeric0\([0, \infty)\)
interaction_constraintsuntyped--
use_quantized_gradlogicalTRUETRUE, FALSE-
num_grad_quant_binsinteger4\((-\infty, \infty)\)
quant_train_renew_leaflogicalFALSETRUE, FALSE-
stochastic_roundinglogicalTRUETRUE, FALSE-
serializablelogicalTRUETRUE, FALSE-
max_bininteger255\([2, \infty)\)
max_bin_by_featureuntypedNULL-
min_data_in_bininteger3\([1, \infty)\)
bin_construct_sample_cntinteger200000\([1, \infty)\)
data_random_seedinteger1\((-\infty, \infty)\)
is_enable_sparselogicalTRUETRUE, FALSE-
enable_bundlelogicalTRUETRUE, FALSE-
use_missinglogicalTRUETRUE, FALSE-
zero_as_missinglogicalFALSETRUE, FALSE-
feature_pre_filterlogicalTRUETRUE, FALSE-
pre_partitionlogicalFALSETRUE, FALSE-
two_roundlogicalFALSETRUE, FALSE-
forcedbins_filenameuntyped""-
is_unbalancelogicalFALSETRUE, FALSE-
scale_pos_weightnumeric1\([0, \infty)\)
sigmoidnumeric1\([0, \infty)\)
boost_from_averagelogicalTRUETRUE, FALSE-
eval_atuntyped1:5-
multi_error_top_kinteger1\([1, \infty)\)
auc_mu_weightsuntypedNULL-
num_machinesinteger1\([1, \infty)\)
local_listen_portinteger12400\([1, \infty)\)
time_outinteger120\([1, \infty)\)
machinesuntyped""-
gpu_platform_idinteger-1\((-\infty, \infty)\)
gpu_device_idinteger-1\((-\infty, \infty)\)
gpu_device_id_listuntypedNULL-
gpu_use_dplogicalFALSETRUE, FALSE-
num_gpuinteger1\([1, \infty)\)
start_iteration_predictinteger0\((-\infty, \infty)\)
num_iteration_predictinteger-1\((-\infty, \infty)\)
pred_early_stoplogicalFALSETRUE, FALSE-
pred_early_stop_freqinteger10\((-\infty, \infty)\)
pred_early_stop_marginnumeric10\((-\infty, \infty)\)
num_iterationsinteger100\([1, \infty)\)
early_stopping_roundsinteger-\([1, \infty)\)
early_stopping_min_deltanumeric-\([0, \infty)\)
first_metric_onlylogicalFALSETRUE, FALSE-

Initial parameter values

  • num_threads:

    • Actual default: 0L

    • Initial value: 1L

    • Reason for change: Prevents accidental conflicts with future.

  • verbose:

    • Actual default: 1L

    • Initial value: -1L

    • Reason for change: Prevents accidental conflicts with mlr messaging system.

  • objective:

    • Depends on the task: if binary classification, then this parameter is set to "binary", otherwise "multiclass" and cannot be changed.

Custom mlr3 parameters

  • num_class: This parameter is automatically inferred for multiclass tasks and does not have to be set.

Early Stopping and Validation

Early stopping can be used to find the optimal number of boosting rounds. Set early_stopping_rounds to an integer value to monitor the performance of the model on the validation set while training. For information on how to configure the validation set, see the Validation section of mlr3::Learner. The internal validation measure can be set the eval parameter which should be a list of mlr3::Measures, functions, or strings for the internal lightgbm measures. If first_metric_only = FALSE (default), the learner stops when any metric fails to improve.

References

Ke, Guolin, Meng, Qi, Finley, Thomas, Wang, Taifeng, Chen, Wei, Ma, Weidong, Ye, Qiwei, Liu, Tie-Yan (2017). “Lightgbm: A highly efficient gradient boosting decision tree.” Advances in neural information processing systems, 30.

See also

Author

kapsner

Super classes

mlr3::Learner -> mlr3::LearnerClassif -> LearnerClassifLightGBM

Active bindings

internal_valid_scores

The validation scores for all metrics at the best iteration (model$best_iter), which is also the iteration that LightGBM predicts with. Extracted from model$record_evals.

best_valid_scores

The validation scores for all metrics at the best iteration (model$best_iter), i.e. the iteration that is also reported via $internal_tuned_values. Because LightGBM also predicts with the best iteration, these are identical to $internal_valid_scores whenever early stopping is activated. If early stopping is not activated, no best iteration is tracked and this is an empty list.

internal_tuned_values

Returns the early stopped iterations if early_stopping_rounds was set during training.

validate

How to construct the internal validation data. This parameter can be either NULL, a ratio, "test", or "predefined".

Methods

Public methods

Inherited methods

LearnerClassifLightGBM$new()

Creates a new instance of this R6 class.

Usage

LearnerClassifLightGBM$new()

LearnerClassifLightGBM$importance()

The importance scores are extracted from lbg.importance.

Usage

LearnerClassifLightGBM$importance()

Returns

Named numeric().

LearnerClassifLightGBM$clone()

The objects of this class are cloneable with this method.

Usage

LearnerClassifLightGBM$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

# Define the Learner
learner = lrn("classif.lightgbm")
print(learner)
#> 
#> ── <LearnerClassifLightGBM> (classif.lightgbm): Gradient Boosting ──────────────
#> • Model: -
#> • Parameters: verbose=-1, num_threads=1
#> • Validate: NULL
#> • Packages: mlr3, mlr3extralearners, and lightgbm
#> • Predict Types: response and [prob]
#> • Feature Types: logical, integer, numeric, and factor
#> • Encapsulation: none (fallback: -)
#> • Properties: hotstart_forward, importance, internal_tuning, missings,
#> multiclass, twoclass, validation, and weights
#> • Other settings: use_weights = 'use', predict_raw = 'FALSE'

# Define a Task
task = tsk("sonar")

# Create train and test set
ids = partition(task)

# Train the learner on the training ids
learner$train(task, row_ids = ids$train)

print(learner$model)
#> LightGBM Model (100 trees)
#> Objective: binary
#> Fitted to dataset with 60 columns
print(learner$importance())
#>          V11          V12          V20           V1          V36          V27 
#> 1.597416e-01 6.220373e-02 5.387648e-02 5.324806e-02 5.152704e-02 4.608114e-02 
#>          V45          V47          V49          V37          V43          V48 
#> 4.389285e-02 4.355484e-02 3.614121e-02 3.103091e-02 3.096447e-02 2.854371e-02 
#>          V21          V31           V4           V5          V28          V16 
#> 2.670852e-02 2.543275e-02 2.294021e-02 1.968101e-02 1.954525e-02 1.892431e-02 
#>          V51          V29          V39          V54          V46          V60 
#> 1.794270e-02 1.723847e-02 1.677088e-02 1.613613e-02 1.579231e-02 1.437862e-02 
#>          V52          V18           V9          V23          V58          V40 
#> 1.429881e-02 1.333639e-02 1.184254e-02 1.143568e-02 9.813536e-03 9.415549e-03 
#>          V55          V59          V10          V34          V22          V42 
#> 8.763068e-03 6.637325e-03 6.507306e-03 4.957727e-03 4.670226e-03 4.075838e-03 
#>           V8          V25          V30          V44          V32          V38 
#> 3.752850e-03 3.706970e-03 3.549155e-03 2.201324e-03 2.086790e-03 1.318736e-03 
#>          V24          V26          V53          V15          V19          V35 
#> 1.287205e-03 1.082947e-03 7.872261e-04 5.527740e-04 5.244992e-04 4.436889e-04 
#>           V7          V50          V57          V33          V17 
#> 2.290154e-04 1.827998e-04 1.328248e-04 7.070691e-05 3.925987e-05 

# Make predictions for the test rows
predictions = learner$predict(task, row_ids = ids$test)

# Score the predictions
predictions$score()
#> classif.ce 
#>  0.1449275