Implementation:Scikit learn Scikit learn LossFunction

Knowledge Sources	Scikit_learn Scikit-learn Docs
Domains	Machine Learning, Loss Functions, Optimization
Last Updated	2026-02-08 15:00 GMT

Overview

Concrete module providing loss function classes for fitting regression and classification models in scikit-learn.

Description

The _loss/loss module provides a common private framework for loss functions used across scikit-learn estimators including LogisticRegression, PoissonRegressor, GammaRegressor, TweedieRegressor, and gradient boosting models. The BaseLoss class uses composition with Cython loss implementations and link functions. Specific loss classes include HalfSquaredError, HalfBinomialLoss, HalfPoissonLoss, HalfGammaLoss, HalfMultinomialLoss, and more.

Usage

These loss classes are used internally by scikit-learn estimators. They provide gradient, hessian, and loss computation with support for sample weights. Use them when implementing custom gradient-based estimators that need standardized loss computation.

Code Reference

Source Location

• Repository: scikit-learn
• File: sklearn/_loss/loss.py

Signature

class BaseLoss:
    # Conventions:
    #   y_true.shape = sample_weight.shape = (n_samples,)
    #   y_pred.shape = raw_prediction.shape = (n_samples,)
    def __init__(self, closs, link, n_classes=None):
        ...
    def loss(self, y_true, raw_prediction, sample_weight=None, ...):
        ...
    def gradient(self, y_true, raw_prediction, sample_weight=None, ...):
        ...
    def gradient_hessian(self, y_true, raw_prediction, sample_weight=None, ...):
        ...

class HalfSquaredError(BaseLoss):
    ...

class HalfBinomialLoss(BaseLoss):
    ...

class HalfMultinomialLoss(BaseLoss):
    ...

class HalfPoissonLoss(BaseLoss):
    ...

class HalfGammaLoss(BaseLoss):
    ...

Import

from sklearn._loss.loss import BaseLoss, HalfSquaredError, HalfBinomialLoss

I/O Contract

Inputs

Name	Type	Required	Description
y_true	ndarray of shape (n_samples,)	Yes	True target values
raw_prediction	ndarray of shape (n_samples,)	Yes	Raw prediction values (before link function)
sample_weight	ndarray of shape (n_samples,)	No	Per-sample weights

Outputs

Name	Type	Description
loss	float	Aggregated loss value
gradient	ndarray of shape (n_samples,)	Gradient of the loss w.r.t. raw_prediction
hessian	ndarray of shape (n_samples,)	Hessian (second derivative) of the loss

Usage Examples

Basic Usage

import numpy as np
from sklearn._loss.loss import HalfSquaredError

loss_fn = HalfSquaredError()
y_true = np.array([1.0, 2.0, 3.0])
raw_prediction = np.array([1.1, 1.9, 3.2])

loss_value = loss_fn.loss(y_true, raw_prediction)
gradient = loss_fn.gradient(y_true, raw_prediction)

Related Pages

• Principle:Scikit_learn_Scikit_learn_Generalized_Linear_Models