Implementation:Online ml River Tree Nodes SGT

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

Overview

Leaf node implementation for Stochastic Gradient Trees (SGT) that handles gradient and hessian information for gradient boosting.

Description

SGTLeaf is a specialized leaf node for Stochastic Gradient Trees that stores gradient and hessian statistics rather than raw target values. It uses feature quantizers (dynamic or static) to discretize numerical features and maintains split statistics for both categorical and numerical features. The leaf evaluates splits based on delta loss metrics and can update its prediction or split into branches based on gradient information.

Usage

Use SGTLeaf when building Stochastic Gradient Trees for regression or classification tasks using gradient boosting. The tree handles target transformation and encoding, while leaves manage gradient statistics.

Code Reference

Source Location

• Repository: Online_ml_River
• File: river/tree/nodes/sgt_nodes.py

Signature

class SGTLeaf(Leaf):
    def __init__(self, prediction: float = 0.0, depth: int = 0, split_params: dict | None = None):
        ...

    def update(self, x: dict, gh: GradHess, sgt, w: float = 1.0):
        ...

    def prediction(self) -> float:
        ...

    def find_best_split(self, sgt) -> BranchFactory:
        ...

    def apply_split(self, split, p_node, p_branch, sgt):
        ...

    @property
    def total_weight(self) -> float:
        ...

    @staticmethod
    def delta_prediction(gh: GradHess, lambda_value: float):
        ...

Import

from river.tree.nodes.sgt_nodes import SGTLeaf

I/O Contract

Input	Type	Description
prediction	float	Initial prediction value
depth	int	Node depth in tree
x	dict	Feature dictionary
gh	GradHess	Gradient and hessian pair
w	float	Sample weight

Output	Type	Description
prediction	float	Current prediction value
best_split	BranchFactory	Best split candidate including null split
delta_pred	float	Prediction update value

Usage Examples

from river.tree.nodes.sgt_nodes import SGTLeaf
from river.tree.utils import GradHess

# Create SGT leaf
leaf = SGTLeaf(
    prediction=0.5,
    depth=1,
    split_params={}
)

# Update with gradient/hessian
x = {'feature1': 0.5, 'feature2': 1.2}
gh = GradHess(gradient=-0.3, hessian=0.8)
leaf.update(x, gh, sgt=None, w=1.0)

# Get prediction
pred = leaf.prediction()  # Returns 0.5

# Find best split
best_split = leaf.find_best_split(sgt=None)

# Calculate delta prediction
delta = SGTLeaf.delta_prediction(gh, lambda_value=1.0)

Related Pages

• Environment:Online_ml_River_Python_Runtime_Environment