LIGHTGRADBOOST Procedure
PROC LIGHTGRADBOOST Statement
PROC LIGHTGRADBOOST <options>;
The PROC LIGHTGRADBOOST statement invokes the procedure. Table 2 summarizes the options in this statement.
Table 2: PROC LIGHTGRADBOOST Statement Options
| Option | Description |
|---|---|
| Basic Options | |
| BAGGINGFRACTION= | Randomly selects a portion of the observations without resampling |
| BAGGINGFREQUENCY= | Specifies the frequency of bagging |
| BOOSTING= | Specifies the type of boosting |
| DATA= | Specifies the input data table |
| DETERMINISTIC | Ensures stable results when the same data and same parameters are used |
| INPUTFRACTION= | Randomly selects a subset of features on each iteration or tree. If you set it to 0.8, the LightGBM algorithm selects 80% of features before training each tree. |
| LASSO= | Specifies the L1-norm regularization parameter |
| LEAFSIZE= | Specifies the minimum number of observations per leaf |
| LEARNINGRATE= | Specifies the learning rate for each tree |
| MAXCATEGORIES= | Limits the number of split points considered for categorical features |
| MAXDEPTH= | Limits the maximum depth of the tree model. A value less than or equal to 0 means no limit. |
| MAXITERS= | Specifies the number of iterations for the boosting |
| MONOTONELIST= | Specifies constraints of monotone features |
| NTHREADS= | Specifies the number of threads to use in the computation |
| NUMBIN= | Specifies the number of bins to use for numeric variables in calculating the decision tree |
| OBJECTIVE= | Specifies the type of objective |
| RIDGE= | Specifies the L2-norm regularization parameter |
| SEED= | Specifies the random number seed to use for model building |
| USEMISSING | Specifies the use of special handling of missing values |
| VALIDATIONDATA= | Specifies the validation data table |
You can specify the following options:
- BAGGINGFRACTION=number
-
randomly selects a portion of the observations without resampling. If number is nonzero, the LightGBM algorithm enables bagging.
By default, BAGGINGFRACTION=1.0. You can tune this value by using the AUTOTUNE statement.
- BAGGINGFREQUENCY=number
-
specifies the frequency of bagging. If number is greater than 0, the LightGBM algorithm performs bagging. If number is n, the LightGBM algorithm randomly selects a value (
) to use for the next n iterations. By default, BAGGINGFREQUENCY=0. You can tune this value by using the AUTOTUNE statement.
- BOOSTING=DART | GBDT | GOSS | RF
-
specifies the type of boosting to use. You can specify the following values:
- DART
uses dropouts meet multiple additive regression trees.
- GBDT
uses the traditional gradient boosting decision tree.
- GOSS
uses gradient-based one-side sampling.
- RF
uses the random forest. You must specify the BAGGINGFRACTION= and BAGGINGFREQUENCY= options in order to use this boosting type.
By default, BOOSTING=GBDT.
- DATA=libref.data-table
-
names the input data table for PROC LIGHTGRADBOOST to use. The default is the most recently created data table. libref.data-table is a two-level name, where
- libref
refers to a collection of information that is defined in the LIBNAME statement and includes the
library, which includes a path to the data, and a session identifier, which defaults to the active session but which can be explicitly defined in the LIBNAME statement. For more information about libref, see the section Using CAS Sessions and CAS Engine Librefs.- data-table
specifies the name of the input data table.
- DETERMINISTIC
ensures stable results when you use the same data and the same parameters.
-
INPUTFRACTION=number
FEATUREFRACTION=number -
randomly selects a subset of features on each iteration or tree. If number is 0.8, the LightGBM algorithm selects 80% of features before training each tree.
By default, INPUTFRACTION=1.0. You can tune this value by using the AUTOTUNE statement.
-
LEAFSIZE=number
MINLEAFSIZE=number -
specifies the minimum number of observations that each child of a split must contain in the training data table in order for the split to be considered.
By default, LEAFSIZE=20. You can tune this value by using the AUTOTUNE statement.
- LEARNINGRATE=number
-
specifies the learning rate for each tree, where number must be greater than 0.
By default, LEARNINGRATE=0.1. You can tune this value by using the AUTOTUNE statement.
- MAXCATEGORIES=number
limits the number of split points to be considered for categorical features. By default, MAXCATEGORIES=32. This value must be greater than 0.
- MAXDEPTH=number
-
limits the maximum depth of the tree model. A number less than or equal to 0 means no limit.
You can tune this value by using the AUTOTUNE statement.
-
MAXITERS=number
NUMITERS=number -
specifies the maximum number of iterations for the boosting.
By default, MAXITERS=100. You can tune this value by using the AUTOTUNE statement.
-
MONOTONELIST=(–1 0 1)
MONOTONECONSTRAINTS=(–1 0 1) specifies series of constraints of monotone features, where 1 represents increasing constraints, –1 represents decreasing constraints, and 0 represents no constraints.
- NTHREADS=number
specifies the number of threads to use in the computation. The default number of threads is calculated for best performance according to the data.
-
NUMBIN=number
NBINS=number
INTERVALBINS=number -
specifies the number of bins to use for binning the interval input variables. PROC LIGHTGRADBOOST bins continuous predictors to a fixed bin size. This option controls the number of bins and thus also the size of the bins.
By default, NUMBIN=255. You can tune this value by using the AUTOTUNE statement.
- OBJECTIVE=function
-
specifies the objective function to use. You can specify the following values:
- BINARY
uses binary classification.
- CROSS_ENTROPY
uses cross-entropy.
- CROSS_ENTROPY_LAMBDA
uses an alternative parameterization of cross-entropy.
- FAIR
uses fair loss.
- GAMMA
uses gamma regression.
- HUBER
uses Huber regression.
- MAPE
uses mean absolute percentage error regression.
- MULTICLASS
uses multiclass classification.
- MULTICLASSOVA
uses one-versus-all multiclass classification.
- POISSON
uses Poisson regression.
- QUANTILE
uses quantile regression.
- REGRESSION
uses regression.
- REGRESSION_L1
uses L1 regression.
- TWEEDIE
uses Tweedie regression.
By default, OBJECTIVE=REGRESSION.
-
REGL1=number
L1=number
LASSO=number -
specifies the L1-norm regularization parameter, where number must be nonnegative.
By default, REGL1=0. You can tune this value by using the AUTOTUNE statement.
-
REGL2=number
L2=number
RIDGE=number -
specifies the L2-norm regularization parameter on prediction. The value of number must be nonnegative.
By default, REGL2=1. You can tune this value by using the AUTOTUNE statement.
- SEED=number
specifies the initial seed for random number generation for model building.
- USEMISSING
specifies the use of special handling of missing values. You can disable this option by omitting it. The LightGBM model uses not a number (NaN) values to represent missing values by default.
-
VALIDATIONDATA=data
VALIDDATA=data specifies that validation data be used. You can use this option to avoid overfitting the training data, combined with parameters such as early stopping rounds.