Module drift 

Data drift and distribution shift detection

This module provides multivariate drift detection for monitoring feature distributions over time between a reference dataset and a current (production) dataset.

Methods

Method	Description
DriftMethod::KolmogorovSmirnov	2-sample KS test per feature with asymptotic p-value
DriftMethod::PopulationStabilityIndex	Binning-based PSI score
DriftMethod::Wasserstein	W1 (Earth mover’s) distance per feature
DriftMethod::MaximumMeanDiscrepancy	MMD² with RBF kernel (multivariate)

Example

use scirs2_transform::drift::{DriftDetector, DriftDetectorConfig, DriftMethod};
use scirs2_core::ndarray::Array2;

let reference = Array2::<f64>::zeros((100, 3));
let config = DriftDetectorConfig::default();
let detector = DriftDetector::fit(&reference, config);

let current = Array2::<f64>::zeros((80, 3));
let report = detector.detect(&current).expect("detection should succeed");
assert!(!report.drifted, "identical distributions should not drift");

Structs

DriftDetector

Multivariate drift detector that compares a reference distribution against a current (test) distribution feature by feature.

DriftDetectorConfig

Configuration for a DriftDetector.

DriftReport

Report produced by DriftDetector::detect.

Enums

DriftMethod

Method used for drift detection.

Functions

ks_test

Compute the 2-sample KS statistic and asymptotic p-value for two 1D samples.

mmd_rbf

Compute MMD² with RBF kernel for two 1D sample arrays.

psi

Compute PSI between a reference and current 1D distribution.

wasserstein_distance_1d

Compute the W1 (Wasserstein-1) distance between two 1D empirical distributions.