holisticai.bias.mitigation.MLDebiaser

class holisticai.bias.mitigation.MLDebiaser(*args, **kargs)

MLDebiaser postprocessing [1] debias predictions w.r.t. the sensitive class in each demographic group. This procedure takes as input a vector y and solves the optimization problem subject to the statistical parity constraint. This bias mitigation can be used for classification (binary and multiclass).

Parameters

gammafloat

The regularization parameter.

epsfloat

The tolerance for the convergence of the optimization algorithm.

etafloat

The step size for the optimization algorithm.

sgd_stepsint

The number of steps for the stochastic gradient descent optimization.

full_gradient_epochsint

The number of epochs for the full gradient optimization.

batch_sizeint

The batch size for the optimization algorithm.

max_iterint

The maximum number of iterations for the optimization algorithm.

verbosebool

The verbosity of the optimization algorithm.

Examples

>>> from holisticai.bias.mitigation import MLDebiaser
>>> mitigator = MLDebiaser()
>>> mitigator.fit(y_proba, group_a, group_b)
>>> test_data_transformed = mitigator.predict(y_proba, group_a, group_b)

References

[1]

Alabdulmohsin, Ibrahim M., and Mario Lucic. “A near-optimal algorithm for debiasing trained machine learning models.” Advances in Neural Information Processing Systems 34 (2021): 8072-8084.

fit(y_proba: ndarray, group_a: ndarray, group_b: ndarray)

Compute parameters for calibrated equalized odds.

Description

Compute parameters for calibrated equalized odds algorithm.

Parameters

y_predarray-like

Predicted vector (num_examples,).

group_aarray-like

Group membership vector (binary)

group_barray-like

Group membership vector (binary)

Returns

Self

fit_transform(y_proba: ndarray, group_a: ndarray, group_b: ndarray)

Fit and transform

Description

Fit and transform

Parameters

y_probaarray-like

Predicted vector (num_examples,).

group_aarray-like

Group membership vector (binary)

group_barray-like

Group membership vector (binary)

Returns

dictionnary with new predictions

transform(y_proba: ndarray, group_a: ndarray, group_b: ndarray)

Apply transform function to predictions and likelihoods

Description

Use a fitted probability to change the output label and invert the likelihood

Parameters

y_probaarray-like

Predicted probability matrix (nb_examples, nb_classes)

group_aarray-like

Group membership vector (binary)

group_barray-like

Group membership vector (binary)

thresholdfloat

float value to discriminate between 0 and 1

Returns

dict

A dictionnary with new predictions