module change_detection.erics

ERICS Change Detection Method.

The ERICS (Effective and Robust Identification of Concept Shift) change detector was proposed by:

[1] HAUG, Johannes; KASNECI, Gjergji. Learning Parameter Distributions to Detect Concept Drift in Data Streams. In: 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021. S. 9452-9459. URL: https://arxiv.org/pdf/2010.09388.pdf

The original source code can be obtained here.

This module provides the ERICS implementation with a Probit base model for binary classification. The update rules for the Probit model are adopted from:

[2] HAUG, Johannes, et al. Leveraging model inherent variable importance for stable online feature selection. In: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020. S. 1478-1502. URL: https://dl.acm.org/doi/abs/10.1145/3394486.3403200

Copyright (C) 2022 Johannes Haug.

class ERICS

ERICS Change Detector.

method ERICS.__init__

__init__(
    n_param: int,
    window_mvg_average: int = 50,
    window_drift_detect: int = 50,
    beta: float = 0.0001,
    init_mu: int = 0,
    init_sigma: int = 1,
    epochs: int = 10,
    lr_mu: float = 0.01,
    lr_sigma: float = 0.01,
    reset_after_drift: bool = False
)

Inits the change detector.

Args:

  • n_param: The total number of parameters in the Probit model. This corresponds to the number of input features.
  • window_mvg_average: The window size for the moving average aggregation of KL divergence measures between the model parameter distributions.
  • window_drift_detect: The window size that is used to compute the pairwise differences between subsequent measures of the moving average. This window and information is used for the change detection.
  • beta: The scaling rate for the automatic update of the alpha-threshold, which is in turn applied to the window_drift_detect to detect concept drift.
  • init_mu: The initial mean of the model parameter distributions.
  • init_sigma: The initial variance of the model parameter distributions.
  • epochs: The number of epochs per optimization iteration of the parameter distributions.
  • lr_mu: The learning rate for the gradient updates of the means.
  • lr_sigma: The learning rate for the gradient updates of the variances.
  • reset_after_drift: A boolean indicating if the change detector will be reset after a drift was detected. This is set to False for ERICS, as this change detector does not need to be reset.

method ERICS.detect_change

detect_change() → bool

Detects global concept drift.

Returns:

  • bool: True, if a concept drift was detected, False otherwise.

method ERICS.detect_partial_change

detect_partial_change() → Tuple[bool, list]

Detects partial concept drift.

Returns:

  • bool: True, if at least one partial concept drift was detected, False otherwise.
  • list: Indices (i.e. relative positions in the feature vector) of input features with detected partial drift.

method ERICS.detect_warning_zone

detect_warning_zone() → bool

Detects a warning zone.

Notes:

ERICS does not raise warnings.

method ERICS.partial_fit

partial_fit(
    X: Union[numpy._array_like._SupportsArray[numpy.dtype], numpy._nested_sequence._NestedSequence[numpy._array_like._SupportsArray[numpy.dtype]], bool, int, float, complex, str, bytes, numpy._nested_sequence._NestedSequence[Union[bool, int, float, complex, str, bytes]]],
    y: Union[numpy._array_like._SupportsArray[numpy.dtype], numpy._nested_sequence._NestedSequence[numpy._array_like._SupportsArray[numpy.dtype]], bool, int, float, complex, str, bytes, numpy._nested_sequence._NestedSequence[Union[bool, int, float, complex, str, bytes]]]
)

Updates the change detector.

Args:

  • X: Batch of observations.
  • y: Batch of labels.

method ERICS.reset

reset()

Resets the change detector.

Notes:

ERICS need not be reset after a drift was detected.

This file was automatically generated via lazydocs.