Implementation:Online ml River Optim Losses

Knowledge Sources	Online_ml_River
Domains	Online_Learning, Optimization
Last Updated	2026-02-08 16:00 GMT

Overview

A comprehensive collection of loss functions for online learning, supporting binary classification, multi-class classification, and regression tasks.

Description

The losses module provides a complete set of differentiable loss functions designed for online machine learning. Each loss function implements both the loss calculation and its gradient, supporting both scalar and numpy vector operations. The module includes specialized losses for binary classification (Log, Hinge, BinaryFocalLoss), multi-class classification (CrossEntropy), and regression (Squared, Absolute, Huber, Quantile, Poisson, Cauchy, EpsilonInsensitiveHinge). All losses inherit from base Loss classes (BinaryLoss, MultiClassLoss, RegressionLoss) that provide appropriate mean functions for their respective tasks.

Usage

Import loss functions from river.optim.losses to use as objective functions in optimizers and linear models. Each loss provides __call__ for computing loss values and gradient methods for optimization.

Code Reference

Source Location

• Repository: Online_ml_River
• File: river/optim/losses.py

Signature

# Base Classes
class BinaryLoss(Loss):
    def mean_func(self, y_pred): ...

class MultiClassLoss(Loss):
    def mean_func(self, y_pred): ...

class RegressionLoss(Loss):
    def mean_func(self, y_pred): ...

# Regression Losses
class Absolute(RegressionLoss):
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Squared(RegressionLoss):
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Huber(RegressionLoss):
    def __init__(self, epsilon=0.1): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Quantile(RegressionLoss):
    def __init__(self, alpha=0.5): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Cauchy(RegressionLoss):
    def __init__(self, C=80): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Poisson(RegressionLoss):
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class EpsilonInsensitiveHinge(RegressionLoss):
    def __init__(self, eps=0.1): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

# Binary Classification Losses
class Log(BinaryLoss):
    def __init__(self, weight_pos=1.0, weight_neg=1.0): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class Hinge(BinaryLoss):
    def __init__(self, threshold=1.0): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

class BinaryFocalLoss(BinaryLoss):
    def __init__(self, gamma=2, beta=1): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

# Multi-Class Classification Loss
class CrossEntropy(MultiClassLoss):
    def __init__(self, class_weight: dict | None = None): ...
    def __call__(self, y_true, y_pred): ...
    def gradient(self, y_true, y_pred): ...

Import

from river import optim

# Access specific losses
loss = optim.losses.Squared()
loss = optim.losses.Log()
loss = optim.losses.CrossEntropy()

I/O Contract

Inputs

Name	Type	Required	Description
y_true	float or dict	Yes	True target value (scalar for regression/binary, dict for multi-class)
y_pred	float or dict	Yes	Predicted value (scalar for regression/binary, dict for multi-class)

Outputs

Name	Type	Description
loss_value	float	Computed loss value from __call__ method
gradient	float, int, or dict	Gradient with respect to prediction from gradient method

Usage Examples

from river import optim

# Squared Loss (L2)
loss = optim.losses.Squared()
print(loss(-4, 5))  # 81
print(loss.gradient(-4, 5))  # 18

# Absolute Loss (L1)
loss = optim.losses.Absolute()
print(loss(-42, 42))  # 84
print(loss.gradient(1, 2))  # 1

# Log Loss for Binary Classification
loss = optim.losses.Log()
# Expects logit (raw) predictions
y_true = 1
y_pred = 2.5  # logit
print(loss(y_true, y_pred))
print(loss.gradient(y_true, y_pred))

# Cross Entropy for Multi-Class
loss = optim.losses.CrossEntropy()
y_true = 2
y_pred = {0: 0.2, 1: 0.3, 2: 0.5}
print(loss(y_true, y_pred))  # -log(0.5)
print(loss.gradient(y_true, y_pred))  # dict with gradients

# Quantile Loss
loss = optim.losses.Quantile(alpha=0.5)  # Median
print(loss(1, 3))  # 1.0
print(loss.gradient(1, 3))  # 0.5

# Huber Loss (robust to outliers)
loss = optim.losses.Huber(epsilon=0.1)
print(loss(0, 0.05))  # Small error: quadratic
print(loss(0, 2.0))   # Large error: linear

Related Pages

• Environment:Online_ml_River_Python_Runtime_Environment