Implementation:Online ml River Drift ADWIN

Knowledge Sources	Domains	Last Updated
River River Docs Learning from Time-Changing Data with Adaptive Windowing	Online Machine Learning, Concept Drift Detection, Statistical Testing	2026-02-08 16:00 GMT

Overview

Concrete tool for detecting concept drift in a data stream using the ADWIN adaptive windowing algorithm, which maintains a variable-length window and tests for statistically significant changes in sub-window means.

Description

The drift.ADWIN class implements the ADWIN2 algorithm for change detection in streaming data. It maintains a variable-length window of recent values using a bucket-based compression scheme. At regular intervals (controlled by the clock parameter), it tests all possible partitions of the window for significant mean differences using the Hoeffding bound. When drift is detected, the oldest portion of the window is dropped.

Internally, the heavy computation is delegated to a C extension (adwin_c.AdaptiveWindowing) for performance. The class exposes properties for inspecting the current window state: width, estimation (mean), variance, total, and n_detections.

Usage

Import drift.ADWIN when you need a standalone drift detector to monitor a scalar signal (such as classification error, loss values, or any numeric metric) for distributional changes. It is also the default drift detector used internally by tree.HoeffdingAdaptiveTreeClassifier and forest.ARFClassifier.

Code Reference

Source Location

river/drift/adwin.py:L8-L135

Signature

class ADWIN(DriftDetector):
    def __init__(
        self,
        delta=0.002,
        clock=32,
        max_buckets=5,
        min_window_length=5,
        grace_period=10,
    )

Import

from river import drift

Key Parameters

Parameter	Type	Default	Description
`delta`	float	0.002	Confidence parameter for the drift test. Lower values mean fewer false positives but slower detection
`clock`	int	32	How often to check for drift. 1 = every new data point; higher values speed up processing but increase detection delay
`max_buckets`	int	5	Maximum number of buckets of each size before merging (controls compression granularity)
`min_window_length`	int	5	Minimum sub-window length to consider during drift checks. Smaller values may increase false positives
`grace_period`	int	10	Number of data points to observe before beginning drift detection

Internal Delegation

The class delegates core computation to adwin_c.AdaptiveWindowing, a C extension that handles the bucket-based storage and Hoeffding bound calculations for performance.

I/O Contract

Inputs

Method	Parameter	Type	Description
`update`	x	float	A single numeric value to add to the window (e.g., 0/1 for correct/incorrect classification)

Outputs

Property/Method	Return Type	Description
`drift_detected`	bool	True if drift was detected on the most recent `update` call
`width`	int	Current window size
`estimation`	float	Mean of the values in the current window
`variance`	float	Sample variance within the adaptive window
`total`	float	Sum of all values in the window
`n_detections`	int	Total number of drifts detected over the detector's lifetime

Usage Examples

Basic Drift Detection

import random
from river import drift

rng = random.Random(12345)
adwin = drift.ADWIN()

# Simulate a data stream with a distribution change at index 1000
data_stream = rng.choices([0, 1], k=1000) + rng.choices(range(4, 8), k=1000)

for i, val in enumerate(data_stream):
    adwin.update(val)
    if adwin.drift_detected:
        print(f"Change detected at index {i}, input value: {val}")
# Change detected at index 1023, input value: 4

Monitoring Classification Error

from river import drift

adwin = drift.ADWIN(delta=0.001)

# Feed binary error indicators (0 = correct, 1 = incorrect)
for i, error in enumerate(errors):
    adwin.update(error)
    if adwin.drift_detected:
        print(f"Drift detected at step {i}")
        print(f"Window size: {adwin.width}, Mean error: {adwin.estimation:.4f}")

Tuning Sensitivity

from river import drift

# More sensitive (more false positives, faster detection)
sensitive_adwin = drift.ADWIN(delta=0.01, clock=1)

# Less sensitive (fewer false positives, slower detection)
conservative_adwin = drift.ADWIN(delta=0.0001, clock=64)

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment