Module drift 

Drift detection and adaptation.

This module provides bounded-memory drift detection algorithms, a decoupled action/strategy layer, decay-aware learning utilities, and a DriftAwareModel wrapper that integrates drift detection into the predict → learn loop.

Overview

• Detectors: PageHinkley, Adwin, Kswin — each implements the DriftDetector trait and reports a DriftLevel (None / Warning / Drift).
• Actions: DriftAction describes what to do when drift is detected. DriftStrategy maps a level to an action, keeping detection and response decoupled.
• Decay learning: TimeDecayedMean, LearningRateScheduler, FixedWindowBuffer — utilities for adapting to non-stationary streams.
• Wrapper: DriftAwareModel wraps a model + detector + strategy and automatically responds to drift during learn. It does not auto-reset the model by default.

Quick start

use rill_ml::drift::{DriftAction, DriftLevel, PageHinkley, StaticStrategy};
use rill_ml::drift::{DriftDetector, DriftStrategy};

let mut detector = PageHinkley::default();
let strategy = StaticStrategy::new(
    DriftAction::ReduceConfidence,
    DriftAction::ResetModel,
);

// Feed a stable stream.
for _ in 0..100 {
    detector.update(0.0).unwrap();
}
assert_eq!(detector.level(), DriftLevel::None);

// Introduce a sudden shift.
for _ in 0..50 {
    detector.update(5.0).unwrap();
}
assert!(detector.detected());
let action = strategy.decide(detector.level(), detector.samples_seen());
assert_eq!(action, DriftAction::ResetModel);

Re-exports

pub use action::DriftAction;

pub use action::DriftEvent;

pub use adwin::Adwin;

pub use adwin::AdwinConfig;

pub use aware_model::DriftAwareModel;

pub use decay::FixedWindowBuffer;

pub use decay::LearningRateScheduler;

pub use decay::TimeDecayedMean;

pub use detector::DriftDetector;

pub use detector::DriftLevel;

pub use kswin::Kswin;

pub use kswin::KswinConfig;

pub use page_hinkley::PageHinkley;

pub use page_hinkley::PageHinkleyConfig;

pub use strategy::DriftStrategy;

pub use strategy::StaticStrategy;

Modules

action

Drift action and event types.

adwin

ADWIN (Adaptive Windowing) drift detector.

aware_model

A model wrapper that integrates drift detection into the predict → learn loop.

decay

Decay-aware learning utilities.

detector

Core drift detector trait.

kswin

KSWIN (Kolmogorov-Smirnov Windowing) drift detector.

page_hinkley

Page-Hinkley sequential change detection.

strategy

Drift handling strategies.