Implementation:Online ml River Drift DriftRetrainingClassifier

Knowledge Sources	Domains	Last Updated
River River Docs	Online Machine Learning, Concept Drift, Meta-Learning	2026-02-08 16:00 GMT

Overview

Concrete tool for wrapping any base classifier with automatic drift detection and model retraining, providing a modular meta-learning approach to concept drift adaptation.

Description

The drift.DriftRetrainingClassifier is a wrapper class that composes a base classifier with a drift detector. At each learn_one call, the wrapper first makes a prediction using the current model, computes a binary error indicator (0 = correct, 1 = incorrect), and feeds it to the drift detector. Depending on the detector's state and the train_in_background setting:

Warning detected + background training enabled: The background model (bkg_model) is trained on the current sample.
Drift detected + background training enabled: The primary model is replaced by the background model, and a fresh background model is cloned.
Drift detected + background training disabled: The primary model is cloned (reset to a fresh state).

The class inherits from both base.Wrapper and base.Classifier, making it compatible with River's evaluation and pipeline infrastructure.

Usage

Import drift.DriftRetrainingClassifier when you want to add drift adaptation to any River classifier (e.g., HoeffdingTreeClassifier, LogisticRegression, GaussianNB) without modifying the classifier itself.

Code Reference

Source Location

river/drift/retrain.py:L6-L100

Signature

class DriftRetrainingClassifier(base.Wrapper, base.Classifier):
    def __init__(
        self,
        model: base.Classifier,
        drift_detector: base.DriftAndWarningDetector
            | base.BinaryDriftAndWarningDetector
            | None = None,
        train_in_background: bool = True,
    )

Import

from river import drift

Key Parameters

Parameter	Type	Default	Description
`model`	base.Classifier	(required)	The base classifier to wrap. Will be cloned upon drift detection
`drift_detector`	DriftAndWarningDetector or None	None (defaults to `drift.binary.DDM()`)	The drift detector for monitoring prediction errors. Must support both warning and drift signals when `train_in_background=True`
`train_in_background`	bool	True	If True, train a background model during warning phase and swap on drift. If False, simply reset the model on drift

I/O Contract

Inputs

Method	Parameter	Type	Description
`learn_one`	x	dict	Feature dictionary
`learn_one`	y	ClfTarget	Target class label
`learn_one`	**kwargs	dict	Additional keyword arguments passed to the base model
`predict_proba_one`	x	dict	Feature dictionary
`predict_proba_one`	**kwargs	dict	Additional keyword arguments passed to the base model

Outputs

Method	Return Type	Description
`predict_proba_one(x)`	dict	Class probability distribution (delegated to the wrapped model)
`predict_one(x)`	ClfTarget	Predicted class label (delegated to the wrapped model)

Internal State

Attribute	Type	Description
`model`	base.Classifier	The current primary model (may be swapped on drift)
`bkg_model`	base.Classifier	The background model (exists only when `train_in_background=True`)
`drift_detector`	DriftDetector	The drift detector instance; inspect `drift_detector.drift_detected` for state

Usage Examples

Basic Usage with DDM Drift Detector

from river import datasets, evaluate, drift, metrics, tree

dataset = datasets.Elec2().take(3000)

model = drift.DriftRetrainingClassifier(
    model=tree.HoeffdingTreeClassifier(),
    drift_detector=drift.binary.DDM()
)

metric = metrics.Accuracy()
evaluate.progressive_val_score(dataset, model, metric)
# Accuracy: 86.46%

Without Background Training

from river import drift, tree

model = drift.DriftRetrainingClassifier(
    model=tree.HoeffdingTreeClassifier(),
    drift_detector=drift.binary.DDM(),
    train_in_background=False
)

# On drift, the model is immediately reset (no background training)

With Logistic Regression and Pipeline

from river import drift, linear_model, preprocessing

model = drift.DriftRetrainingClassifier(
    model=preprocessing.StandardScaler() | linear_model.LogisticRegression(),
    drift_detector=drift.binary.DDM()
)

Inspecting Drift Detection State

from river import drift, tree, datasets

model = drift.DriftRetrainingClassifier(
    model=tree.HoeffdingTreeClassifier(),
    drift_detector=drift.binary.DDM()
)

n_drifts = 0
for x, y in datasets.Elec2().take(5000):
    model.learn_one(x, y)
    if model.drift_detector.drift_detected:
        n_drifts += 1
        print(f"Drift #{n_drifts} detected")

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment