Implementation:Online ml River Tree Splitter SGTQuantizer

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

Overview

Feature quantizers for Stochastic Gradient Trees that discretize features using dynamic or static quantization strategies.

Description

This module provides two quantization strategies for SGT. DynamicQuantizer starts with an initial radius and adapts it to the data's standard deviation for new quantizers. StaticQuantizer buffers initial samples to determine fixed quantization bins that are replicated to all new quantizers. Both track gradient and hessian statistics in bins. Dynamic quantization adapts to each feature's scale, while static quantization provides consistent bins across the tree.

Usage

Use DynamicQuantizer when features have varying scales and you want adaptive quantization. Use StaticQuantizer when you want consistent binning based on initial data distribution.

Code Reference

Source Location

• Repository: Online_ml_River
• File: river/tree/splitter/sgt_quantizer.py

Signature

class DynamicQuantizer(Quantizer):
    def __init__(self, radius: float = 0.5, std_prop: float = 0.25):
        ...

    def update(self, x_val, gh: GradHess, w: float):
        ...

    def __len__(self):
        ...

    def __iter__(self):
        ...

    def _get_params(self):
        ...


class StaticQuantizer(Quantizer):
    def __init__(self, n_bins: int = 64, warm_start: int = 100, *, buckets: list | None = None):
        ...

    def update(self, x_val, gh: GradHess, w: float):
        ...

    def __len__(self):
        ...

    def __iter__(self):
        ...

    def _get_params(self):
        ...

Import

from river.tree.splitter.sgt_quantizer import DynamicQuantizer
from river.tree.splitter.sgt_quantizer import StaticQuantizer

I/O Contract

Input	Type	Description
x_val	float	Feature value
gh	GradHess	Gradient and hessian pair
w	float	Sample weight
radius	float	Initial quantization radius (DynamicQuantizer)
n_bins	int	Number of bins (StaticQuantizer)
warm_start	int	Warmup samples (StaticQuantizer)

Output	Type	Description
__len__	int	Number of bins
__iter__	Iterator[tuple]	(threshold, GradHessStats) pairs
_get_params	dict	Parameters for cloning

Usage Examples

from river.tree.splitter.sgt_quantizer import DynamicQuantizer, StaticQuantizer
from river.tree.utils import GradHess

# Dynamic quantizer adapts to data scale
dq = DynamicQuantizer(radius=0.5, std_prop=0.25)

for i in range(100):
    x_val = float(i) / 10
    gh = GradHess(gradient=-0.1, hessian=0.2)
    dq.update(x_val, gh, w=1.0)

print(f"Number of bins: {len(dq)}")

# Iterate over bins
for threshold, ghs in dq:
    print(f"Threshold: {threshold}, Total weight: {ghs.total_weight}")

# Get adapted parameters for new quantizer
params = dq._get_params()

# Static quantizer uses fixed bins
sq = StaticQuantizer(n_bins=64, warm_start=100)

# Update during warmup (fills buffer)
for i in range(100):
    gh = GradHess(gradient=-0.1, hessian=0.2)
    sq.update(float(i), gh, w=1.0)

# After warmup, bins are fixed
print(f"Bins created: {len(sq)}")

Related Pages

• Environment:Online_ml_River_Python_Runtime_Environment