Skip to main content

Module drift

Module drift 

Source
Expand description

Distribution drift detection for model monitoring.

Three techniques:

  • PSI (Population Stability Index): compares binned distributions of current vs training data. Threshold ≥ 0.25 → significant drift.
  • Kolmogorov–Smirnov: two-sample KS test on CDFs.
  • ADWIN (Adaptive Windowing): online algorithm that detects concept drift in a data stream by maintaining adaptive windows.

Structs§

DriftMonitor
Background monitor that periodically checks registered models for data drift.
ModelDriftReport
Complete drift analysis report.

Enums§

DriftAction
Action to take when drift is detected.
DriftDetector
Drift detection strategy.

Type Aliases§

DriftCallback
Callback invoked when drift is detected on a model.