import sys
import warnings
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, ClassVar, TypeVar
import numpy as np
from imblearn.base import BaseSampler
from numpy.typing import ArrayLike, NDArray
from sklearn.base import clone
from sklearn.utils import _safe_indexing
from sklearn.utils._param_validation import HasMethods, StrOptions
from .._types import FloatNDArray, IntNDArray, ParameterConstraint
from ..utils._sklearn_compat import ClassifierTags, Tags, type_of_target # type: ignore
from ._strategies import Strategy
if sys.version_info >= (3, 11):
from typing import Self
else:
from typing_extensions import Self
if TYPE_CHECKING: # pragma: no cover
import pandas as pd
_XT = TypeVar('_XT', NDArray[Any], pd.DataFrame, ArrayLike)
_YT = TypeVar('_YT', NDArray[Any], pd.Series, ArrayLike)
else:
_XT = TypeVar('_XT', NDArray[Any], ArrayLike)
_YT = TypeVar('_YT', NDArray[Any], ArrayLike)
StrategyFn = Callable[[NDArray[Any], NDArray[Any]], int]
def _independent_pairs(y_true: ArrayLike, sensitive_feature: NDArray[Any]) -> int:
"""Determine promotion and demotion pairs so that y is statistically independent of sensitive feature."""
sensitive_indices = np.where(sensitive_feature == 0)[0]
not_sensititive_indices = np.where(sensitive_feature == 1)[0]
n_sensitive = len(sensitive_indices)
n_not_sensitive = len(not_sensititive_indices)
n = n_sensitive + n_not_sensitive
# no swapping needed if one of the groups is empty
if n_sensitive == 0 or n_not_sensitive == 0:
warnings.warn(
'sensitive_feature only contains one class, no relabeling is performed.',
UserWarning,
stacklevel=2,
)
return 0
pos_ratio_sensitive = np.sum(_safe_indexing(y_true, sensitive_indices)) / n_sensitive
pos_ratio_not_sensitive = np.sum(_safe_indexing(y_true, not_sensititive_indices)) / n_not_sensitive
discrimination = pos_ratio_not_sensitive - pos_ratio_sensitive
# number of pairs to swap label
return int(abs(round((discrimination * n_sensitive * n_not_sensitive) / n)))
[docs]
class BiasRelabler(BaseSampler): # type: ignore[misc]
"""
Sampler which relabels instances to remove bias against a subgroup.
Read more in the :ref:`User Guide <bias_mitigation>`.
Parameters
----------
estimator : Estimator instance
Base estimator which is used to determine the number of promotion and demotion pairs.
strategy : {'statistical parity', 'demographic parity'} or Callable, default='statistical parity'
Determines how the group weights are computed.
Group weights determine how many instances to relabel for each combination of target and sensitive_feature.
- ``'statistical_parity'`` or ``'demographic parity'``: \
probability of positive predictions are equal between subgroups of sensitive feature.
- ``Callable``: function which computes the number of labels swaps based on the target and sensitive feature. \
Callable accepts two arguments: \
y_true and sensitive_feature and returns the number of pairs needed to be swapped.
transform_feature : Optional[Callable[[numpy.ndarray], numpy.ndarray]], default=None
Function which transforms sensitive feature before resampling the training data.
The function takes in the sensitive feature in the form of a numpy.ndarray
and outputs the transformed sensitive feature as a numpy.ndarray.
This can be useful if you want to transform a continuous variable to a binary variable at fit time.
Attributes
----------
estimator_ : Estimator instance
Fitted estimator.
Examples
--------
.. code-block:: python
import numpy as np
from empulse.samplers import BiasRelabler
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
X, y = make_classification()
high_clv = np.random.randint(0, 2, y.shape)
sampler = BiasRelabler(LogisticRegression())
sampler.fit_resample(X, y, sensitive_feature=high_clv)
Example with passing high-clv indicator through cross-validation:
.. code-block:: python
import numpy as np
from empulse.samplers import BiasRelabler
from imblearn.pipeline import Pipeline
from sklearn import set_config
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
set_config(enable_metadata_routing=True)
X, y = make_classification()
high_clv = np.random.randint(0, 2, y.shape)
pipeline = Pipeline([
('sampler', BiasRelabler(
LogisticRegression()
).set_fit_resample_request(sensitive_feature=True)),
('model', LogisticRegression())
])
cross_val_score(pipeline, X, y, params={'sensitive_feature': high_clv})
Example with passing clv through a grid search and dynamically determining high_clv customer based on training data:
.. code-block:: python
import numpy as np
from empulse.samplers import BiasRelabler
from imblearn.pipeline import Pipeline
from sklearn import set_config
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import GridSearchCV
set_config(enable_metadata_routing=True)
X, y = make_classification()
clv = np.random.rand(y.size)
def to_high_clv(clv: np.ndarray) -> np.ndarray:
return (clv > np.median(clv)).astype(np.int8)
pipeline = Pipeline([
('sampler', BiasRelabler(
LogisticRegression(),
transform_feature=to_high_clv
).set_fit_resample_request(sensitive_feature=True)),
('model', LogisticRegression())
])
param_grid = {'model__C': np.logspace(-5, 2, 10)}
grid_search = GridSearchCV(pipeline, param_grid=param_grid)
grid_search.fit(X, y, sensitive_feature=clv)
References
----------
.. [1] Rahman, S., Janssens, B., & Bogaert, M. (2025).
Profit-driven pre-processing in B2B customer churn modeling using fairness techniques.
Journal of Business Research, 189, 115159. doi:10.1016/j.jbusres.2024.115159
"""
_estimator_type: ClassVar[str] = 'sampler'
_sampling_type: ClassVar[str] = 'bypass'
_parameter_constraints: ClassVar[ParameterConstraint] = {
'estimator': [HasMethods(['fit', 'predict_proba'])],
'strategy': [StrOptions({'statistical parity', 'demographic parity'}), callable],
'transform_feature': [callable, None],
}
_strategy_mapping: ClassVar[dict[Strategy, StrategyFn]] = {
'statistical parity': _independent_pairs,
'demographic parity': _independent_pairs,
}
if TYPE_CHECKING: # pragma: no cover
# BaseEstimator should dynamically generate the method signature at runtime
def set_fit_resample_request(self, sensitive_feature: bool = False) -> Self: # noqa: D102
pass
def __init__(
self,
estimator: Any,
*,
strategy: StrategyFn | Strategy = 'statistical parity',
transform_feature: Callable[[NDArray[Any]], IntNDArray] | None = None,
):
super().__init__()
self.estimator = estimator
self.transform_feature = transform_feature
self.strategy = strategy
def _more_tags(self) -> dict[str, bool]:
return {
'binary_only': True,
'poor_score': True,
}
def __sklearn_tags__(self) -> Tags:
tags = super().__sklearn_tags__()
tags.classifier_tags = ClassifierTags(multi_class=False)
return tags
[docs]
def fit_resample(self, X: _XT, y: _YT, *, sensitive_feature: ArrayLike | None = None) -> tuple[_XT, _YT]:
"""
Fit the estimator and relabel the data according to the strategy.
Parameters
----------
X : 2D array-like, shape=(n_samples, n_features)
y : 1D array-like, shape=(n_samples,)
sensitive_feature : 1D array-like, shape=(n_samples,)
Sensitive feature used to determine the number of promotion and demotion pairs.
Returns
-------
X : 2D array-like, shape=(n_samples, n_features)
Original training data.
y : np.ndarray
Relabeled target values.
"""
X, y = super().fit_resample(X, y, sensitive_feature=sensitive_feature)
X: _XT
y: _YT
return X, y
def _fit_resample(
self, X: NDArray[Any], y: NDArray[Any], *, sensitive_feature: ArrayLike | None = None
) -> tuple[NDArray[Any], NDArray[Any]]:
"""
Fit the estimator and relabel the data according to the strategy.
Parameters
----------
X : 2D array-like, shape=(n_samples, n_features)
y : 1D array-like, shape=(n_samples,)
sensitive_feature : 1D array-like, shape=(n_samples,)
Sensitive feature used to determine the number of promotion and demotion pairs.
Returns
-------
X : 2D array-like, shape=(n_samples, n_features)
Original training data.
y : np.ndarray
Relabeled target values.
"""
sensitive_feature = np.asarray(sensitive_feature)
y_type = type_of_target(y, input_name='y', raise_unknown=True)
if y_type != 'binary':
raise ValueError(f'Only binary classification is supported. The type of the target is {y_type}.')
self.classes_: NDArray[np.int64] = np.unique(y)
if len(self.classes_) == 1:
return X, y
y_binarized = np.where(y == self.classes_[1], 1, 0)
if self.transform_feature is not None:
sensitive_feature = self.transform_feature(sensitive_feature)
self.estimator_ = clone(self.estimator)
self.estimator_.fit(X, y)
y_pred = self.estimator_.predict_proba(X)[:, 1]
strategy = self._strategy_mapping[self.strategy] if isinstance(self.strategy, str) else self.strategy
n_pairs = strategy(y_binarized, sensitive_feature)
if n_pairs <= 0:
return X, np.asarray(y)
sensitive_indices = np.where(sensitive_feature == 0)[0]
non_sensitive = np.where(sensitive_feature == 1)[0]
probas_non_sensitive = y_pred[non_sensitive]
probas_sensitive = y_pred[sensitive_indices]
demotion_candidates = _get_demotion_candidates(probas_non_sensitive, _safe_indexing(y, non_sensitive), n_pairs)
promotion_candidates = _get_promotion_candidates(
probas_sensitive, _safe_indexing(y, sensitive_indices), n_pairs
)
# map promotion and demotion candidates to original indices
indices = np.arange(len(y))
demotion_candidates = indices[non_sensitive][demotion_candidates]
promotion_candidates = indices[sensitive_indices][promotion_candidates]
# relabel the data
if hasattr(y, 'copy'):
relabled_y = y.copy()
elif hasattr(y, 'clone'):
relabled_y = y.clone()
else:
relabled_y = np.copy(y)
if hasattr(relabled_y, 'loc'):
relabled_y.loc[demotion_candidates] = 0
relabled_y.loc[promotion_candidates] = 1
else:
relabled_y[demotion_candidates] = 0
relabled_y[promotion_candidates] = 1
return X, relabled_y
def _get_demotion_candidates(y_pred: FloatNDArray, y_true: FloatNDArray, n_pairs: int) -> FloatNDArray:
"""Return the n_pairs instances with the lowest probability of being positive class label."""
positive_indices = np.where(y_true == 1)[0]
positive_predictions = y_pred[positive_indices]
demotion_candidates: FloatNDArray = positive_indices[np.argsort(positive_predictions)[:n_pairs]]
return demotion_candidates
def _get_promotion_candidates(y_pred: FloatNDArray, y_true: FloatNDArray, n_pairs: int) -> FloatNDArray:
"""Return the n_pairs instances with the lowest probability of being negative class label."""
negative_indices = np.where(y_true == 0)[0]
negative_predictions = y_pred[negative_indices]
promotion_candidates: FloatNDArray = negative_indices[np.argsort(negative_predictions)[-n_pairs:]]
return promotion_candidates
