boosters.Objective#

class boosters.Objective#

Bases: object

Objective (loss) functions for gradient boosting.

Each variant represents a different loss function for training GBDT and GBLinear models. Use the static constructor methods for validation.

Regression:

  • Objective.Squared(): Mean squared error (L2)

  • Objective.Absolute(): Mean absolute error (L1)

  • Objective.Huber(delta): Pseudo-Huber loss (robust)

  • Objective.Pinball(alpha): Quantile regression

  • Objective.Poisson(): Poisson deviance for count data

Classification:

  • Objective.Logistic(): Binary cross-entropy

  • Objective.Hinge(): SVM-style hinge loss

  • Objective.Softmax(n_classes): Multiclass cross-entropy

Note: ranking objectives are not implemented in core yet.

Examples

>>> from boosters import Objective
>>> obj = Objective.squared()  # L2 regression
>>> obj = Objective.logistic()  # Binary classification
>>> obj = Objective.pinball([0.1, 0.5, 0.9])  # Quantile regression
>>> obj = Objective.softmax(10)  # Multiclass classification

Pattern matching: >>> match obj: … case Objective.Squared(): … print(“L2 loss”) … case Objective.Pinball(alpha=a): … print(f”Quantile: {a}”)

Absolute#

alias of PyObjective_Absolute

Hinge#

alias of PyObjective_Hinge

Huber#

alias of PyObjective_Huber

Logistic#

alias of PyObjective_Logistic

Pinball#

alias of PyObjective_Pinball

Poisson#

alias of PyObjective_Poisson

Softmax#

alias of PyObjective_Softmax

Squared#

alias of PyObjective_Squared

static absolute()#

Create absolute error loss (L1).

static hinge()#

Create hinge loss for binary classification.

static huber(delta=1.0)#

Create Huber loss with validation.

static logistic()#

Create logistic loss for binary classification.

static pinball(alpha)#

Create pinball loss with validation.

static poisson()#

Create Poisson loss.

static softmax(n_classes)#

Create softmax loss with validation.

static squared()#

Create squared error loss (L2).