from collections.abc import Callable
from typing import Any, ClassVar, Self
import numpy as np
from numpy.typing import ArrayLike, NDArray
from sklearn.base import BaseEstimator, ClassifierMixin, _fit_context, clone
from sklearn.utils._param_validation import HasMethods, StrOptions
from sklearn.utils.validation import check_is_fitted
from ..._types import FloatArrayLike, FloatNDArray, IntNDArray, ParameterConstraint
from ...samplers import BiasRelabler
from ...samplers._strategies import Strategy
from ...utils._sklearn_compat import Tags, type_of_target, validate_data # type: ignore[attr-defined]
StrategyFn = Callable[[NDArray[Any], NDArray[Any]], int]
[docs]
class BiasRelabelingClassifier(ClassifierMixin, BaseEstimator): # type: ignore[misc]
"""
Classifier which relabels instances during training to remove bias against a subgroup.
Read more in the :ref:`User Guide <bias_mitigation>`.
Parameters
----------
estimator : Estimator instance
Base estimator which is used for fitting and predicting.
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], default=None
Function which transforms sensitive feature before resampling the training data.
Attributes
----------
classes_ : numpy.ndarray, shape=(n_classes,)
Unique classes in the target.
estimator_ : Estimator instance
Fitted base estimator.
Examples
--------
1. Using the `BiasRelabelingClassifier` with a logistic regression model:
.. code-block:: python
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
from empulse.models import BiasRelabelingClassifier
X, y = make_classification()
high_clv = np.random.randint(0, 2, size=X.shape[0])
model = BiasRelabelingClassifier(estimator=LogisticRegression())
model.fit(X, y, sensitive_feature=high_clv)
2. Converting a continuous attribute to a binary attribute:
.. code-block:: python
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
from empulse.models import BiasRelabelingClassifier
X, y = make_classification()
clv = np.random.rand(X.shape[0]) * 100
model = BiasRelabelingClassifier(
estimator=LogisticRegression(),
transform_feature=lambda clv: (clv > np.quantile(clv, 0.8)).astype(int)
)
model.fit(X, y, sensitive_feature=clv)
3. Using a custom strategy function:
.. code-block:: python
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import make_classification
from empulse.models import BiasRelabelingClassifier
X, y = make_classification()
high_clv = np.random.randint(0, 2, size=X.shape[0])
# Simple strategy to swap 2 labels
def strategy(y_true, sensitive_feature):
return 2
model = BiasRelabelingClassifier(
estimator=LogisticRegression(),
strategy=strategy
)
model.fit(X, y, sensitive_feature=high_clv)
4. Passing the sensitive feature in a cross-validation grid search:
.. code-block:: python
import numpy as np
from sklearn import config_context
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from empulse.models import BiasRelabelingClassifier
with config_context(enable_metadata_routing=True):
X, y = make_classification()
high_clv = np.random.randint(0, 2, size=X.shape[0])
param_grid = {'model__estimator__C': [0.1, 1, 10]}
pipeline = Pipeline([
('scaler', StandardScaler()),
('model', BiasRelabelingClassifier(LogisticRegression()).set_fit_request(sensitive_feature=True))
])
search = GridSearchCV(pipeline, param_grid)
search.fit(X, y, sensitive_feature=high_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
"""
_parameter_constraints: ClassVar[ParameterConstraint] = {
'estimator': [HasMethods(['fit', 'predict_proba']), None],
'strategy': [callable, StrOptions({'statistical parity', 'demographic parity'}), None],
'transform_feature': [callable, None],
}
def __init__(
self,
estimator: Any,
*,
strategy: StrategyFn | Strategy = 'statistical parity',
transform_feature: Callable[[NDArray[Any]], IntNDArray] | None = None,
):
self.estimator = estimator
self.strategy = strategy
self.transform_feature = transform_feature
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.multi_class = False
tags.classifier_tags.poor_score = True
return tags
[docs]
@_fit_context(prefer_skip_nested_validation=True) # type: ignore[misc]
def fit(self, X: ArrayLike, y: ArrayLike, *, sensitive_feature: ArrayLike | None = None, **fit_params: Any) -> Self:
"""
Fit the estimator and relabels the instances according to the strategy.
Parameters
----------
X : 2D array-like, shape=(n_samples, n_dim)
Training data.
y : 1D array-like, shape=(n_samples,)
Target values.
sensitive_feature : 1D array-like, shape=(n_samples,), default = None
Sensitive feature used to determine the sample group weights.
fit_params : dict
Additional parameters passed to the estimator's `fit` method.
Returns
-------
self : BiasRelabelingClassifier
"""
X, y = validate_data(self, X, y)
y_type = type_of_target(y, input_name='y', raise_unknown=True)
if y_type != 'binary':
raise ValueError(
f'Unknown label type: Only binary classification is supported. The type of the target is {y_type}.'
)
self.classes_ = np.unique(y)
if len(self.classes_) == 1:
raise ValueError("Classifier can't train when only one class is present.")
if sensitive_feature is None:
self.estimator_ = clone(self.estimator)
self.estimator_.fit(X, y, **fit_params)
return self
sensitive_feature = np.asarray(sensitive_feature)
sampler = BiasRelabler(
estimator=self.estimator, strategy=self.strategy, transform_feature=self.transform_feature
)
X, y = sampler.fit_resample(X, y, sensitive_feature=sensitive_feature)
self.estimator_ = clone(self.estimator)
self.estimator_.fit(X, y, **fit_params)
return self
[docs]
def predict_proba(self, X: FloatArrayLike) -> FloatNDArray:
"""
Predict class probabilities for X.
Parameters
----------
X : 2D numpy.ndarray, shape=(n_samples, n_dim)
Returns
-------
y_pred : 2D numpy.ndarray, shape=(n_samples, n_classes)
Predicted class probabilities.
"""
check_is_fitted(self)
X = validate_data(self, X, reset=False)
y_proba: FloatNDArray = self.estimator_.predict_proba(X)
return y_proba
[docs]
def predict(self, X: FloatArrayLike) -> NDArray[Any]:
"""
Predict class labels for X.
Parameters
----------
X : 2D numpy.ndarray, shape=(n_samples, n_dim)
Returns
-------
y_pred : 1D numpy.ndarray, shape=(n_samples,)
Predicted class labels.
"""
y_proba = self.predict_proba(X)
y_pred: NDArray[Any] = self.classes_[np.argmax(y_proba, axis=1)]
return y_pred
