Implementation:Online ml River Tree Splitter Base

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

Overview

Abstract base classes defining the interface for attribute observers (splitters) and feature quantizers in decision trees.

Description

This module provides two abstract base classes: Splitter and Quantizer. Splitter defines the interface for attribute observers that monitor input features and find optimal split points. They track statistics, estimate probability densities, and suggest splits based on criteria. Quantizer is used by Stochastic Gradient Trees to discretize features by tracking gradient and hessian statistics in bins.

Usage

Do not instantiate these base classes directly. Instead, implement concrete splitters (like GaussianSplitter, EBSTSplitter) or quantizers (like DynamicQuantizer, StaticQuantizer) that fulfill their interface contracts.

Code Reference

Source Location

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

Signature

class Splitter(base.Estimator, abc.ABC):
    @abc.abstractmethod
    def update(self, att_val, target_val: base.typing.Target, w: float) -> None:
        pass

    @abc.abstractmethod
    def cond_proba(self, att_val, target_val: base.typing.ClfTarget) -> float:
        pass

    @abc.abstractmethod
    def best_evaluated_split_suggestion(
        self,
        criterion: SplitCriterion,
        pre_split_dist: list | dict,
        att_idx: base.typing.FeatureName,
        binary_only: bool,
    ) -> BranchFactory:
        pass

    @property
    def is_numeric(self) -> bool:
        return True

    @property
    def is_target_class(self) -> bool:
        return True


class Quantizer(base.Estimator, abc.ABC):
    @abc.abstractmethod
    def __len__(self):
        pass

    @abc.abstractmethod
    def update(self, x_val, gh: GradHess, w: float) -> None:
        pass

    @abc.abstractmethod
    def __iter__(self) -> tuple[float, typing.Iterator[GradHessStats]]:
        pass

Import

from river.tree.splitter.base import Splitter
from river.tree.splitter.base import Quantizer

I/O Contract

Input (Splitter)	Type	Description
att_val	any	Feature value
target_val	Target	Target value (class or numeric)
w	float	Sample weight
criterion	SplitCriterion	Split evaluation criterion

Output (Splitter)	Type	Description
cond_proba	float	target) conditional probability
split_suggestion	BranchFactory	Best split candidate with merit score

Usage Examples

# Splitter is abstract, use concrete implementation
from river.tree.splitter import GaussianSplitter
from river.tree.split_criterion import GiniSplitCriterion

splitter = GaussianSplitter(n_splits=10)

# Update with observations
splitter.update(att_val=5.5, target_val='A', w=1.0)
splitter.update(att_val=6.2, target_val='B', w=1.0)

# Get conditional probability
prob = splitter.cond_proba(5.5, 'A')

# Find best split
criterion = GiniSplitCriterion()
pre_split = {'A': 10, 'B': 8}
suggestion = splitter.best_evaluated_split_suggestion(
    criterion=criterion,
    pre_split_dist=pre_split,
    att_idx='feature1',
    binary_only=True
)

print(f"Split merit: {suggestion.merit}")
print(f"Split threshold: {suggestion.split_info}")

Related Pages

• Environment:Online_ml_River_Python_Runtime_Environment