GradientBoostingClassifier

#include <Skigen/Ensemble>

template <typename Scalar = double>
class Skigen::GradientBoostingClassifier(loss=Loss::LogLoss, learning_rate=0.1, n_estimators=100, subsample=1.0, criterion=CriterionGB::FriedmanMSE, min_samples_split=2, min_samples_leaf=1, min_weight_fraction_leaf=0, max_depth=3, min_impurity_decrease=0, random_state=std::nullopt, verbose=0, max_leaf_nodes=std::nullopt, warm_start=false, validation_fraction=0.1, n_iter_no_change=std::nullopt, tol=1e-4, ccp_alpha=0)

Gradient Boosting for binary classification.

Stage-wise additive log-odds model fit by gradient boosting on the negative log-likelihood (cross-entropy) of a binomial outcome:

$$F_M(x) = F_0 + \eta \sum_{m=1}^M h_m(x), \qquad P(y=1 \mid x) = \sigma(F_M(x)).$$

At each stage the pseudo-residual is the negative gradient of the loss: $r_i = y_i - \sigma(F_{m-1}(x_i))$, where $y_i \in \{0, 1\}$. The initial log-odds $F_0$ is the empirical prior log-ratio $\log(p_+ / p_-)$, matching sklearn's init="zero" default which uses a DummyClassifier(strategy="prior") under the hood.

Mirrors sklearn.ensemble.GradientBoostingClassifier for the binary case.

---

---

Attributes:

• loss : Loss

• learning_rate : Scalar

• n_estimators : int

• max_depth : int

• init : Scalar

• classes : const Eigen::VectorXi

• n_classes : int

• estimators : const std::vector< DecisionTreeRegressor< Scalar > > &

• feature_importances : RowVectorType

• train_score : VectorType

---

Methods

SKIGEN_PARAMS()

---

predict(X)

---

decision_function(X)

Raw additive score $F(x)$ (log-odds). Length n_samples for binary; for multiclass use decision_function_multi.

---

decision_function_multi(X)

Per-class raw scores, shape (n_samples, n_classes). For binary problems returns the two complementary log-odds columns.

---

predict_proba(X)

Probability estimates of shape (n_samples, n_classes). Column ordering matches classes().

---