Skip to content

Detectors

Generally, the detection paradigm of driftbench follows the scheme illustrated below. First, features are extracted from each curve, and then aggregations are performed over the temporal axis (often corresponding to process executions) in the resulting latent space.

Overview

Overview

This module provides a collection of drift detectors, ranging from simple baselines to sophisticated statistical and neural approaches. All detectors inherit from a common Detector interface. Here a short overview:

  • SlidingKSWINDetector: Implements a sliding-window Kolmogorov–Smirnov test to compare distributions of past and recent data windows. The size of the windows as well as the offset in between can be parameterized.
  • MMDDetector: Applies the Maximum Mean Discrepancy (MMD) test to compare two samples using kernel methods (multiscale or RBF).
  • AggregateFeatureAlgorithm: Aggregates temporal data (e.g., by computing the mean across time) and applies another drift detector to the result which can be set via algorithm parameter.
  • ClusterDetector: Applies clustering (either KMeans or Gaussian Mixture Model) to the data and uses distance to clusters or likelihoods as drift scores.
  • AutoencoderDetector: Uses a neural autoencoder to reduce dimensionality of the data, then applies another detector in the latent space. Can retrain on each call or reuse the trained model. Hidden layers, activation functions and learning rate can be set via parameters as well as the downstream detector (via parameter detector).
  • RollingMeanDifferenceDetector: Calculates the maximum value over a rolling mean across time and returns the absolute difference between subsequent steps.
  • RandomGuessDetector: Outputs a cumulative sum of random Gaussian noise. Acts as a simple baseline for drift detection.
  • RollingMeanStandardDeviationDetector: Applies a rolling mean followed by a rolling standard deviation and returns the result as the drift score.
  • AlwaysGuessDriftDetector: A trivial detector that always predicts drift. Useful for debugging or as a degenerate upper bound. Provides a simple baseline.