aif360.algorithms.inprocessing.GerryFairClassifier

class aif360.algorithms.inprocessing.GerryFairClassifier(C=10, printflag=False, heatmapflag=False, heatmap_iter=10, heatmap_path='.', max_iters=10, gamma=0.01, fairness_def='FP', predictor=LinearRegression())

Model is an algorithm for learning classifiers that are fair with respect to rich subgroups.

Rich subgroups are defined by (linear) functions over the sensitive attributes, and fairness notions are statistical: false positive, false negative, and statistical parity rates. This implementation uses a max of two regressions as a cost-sensitive classification oracle, and supports linear regression, support vector machines, decision trees, and kernel regression. For details see:

References

[1]

“Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness.” Michale Kearns, Seth Neel, Aaron Roth, Steven Wu. ICML ‘18.

[2]

“An Empirical Study of Rich Subgroup Fairness for Machine Learning”. Michael Kearns, Seth Neel, Aaron Roth, Steven Wu. FAT ‘19.

Initialize Model Object and set hyperparameters.

Parameters:

• C – Maximum L1 Norm for the Dual Variables (hyperparameter)

• printflag – Print Output Flag

• heatmapflag – Save Heatmaps every heatmap_iter Flag

• heatmap_iter – Save Heatmaps every heatmap_iter

• heatmap_path – Save Heatmaps path

• max_iters – Time Horizon for the fictitious play dynamic.

• gamma – Fairness Approximation Paramater

• fairness_def – Fairness notion, FP, FN, SP.

• errors – see fit()

• fairness_violations – see fit()

• predictor – Hypothesis class for the Learner. Supports LR, SVM, KR, Trees.

Methods

fit	Run Fictitious play to compute the approximately fair classifier.
fit_predict	Train a model on the input and predict the labels.
fit_transform	Train a model on the input and transform the dataset accordingly.
generate_heatmap	Helper Function to generate the heatmap at the current time.
pareto	Assumes Model has FP specified for metric.
predict	Return dataset object where labels are the predictions returned by the fitted model.
print_outputs	Helper function to print outputs at each iteration of fit.
save_heatmap	Helper Function to save the heatmap.
transform	Return a new dataset generated by running this Transformer on the input.

__init__(C=10, printflag=False, heatmapflag=False, heatmap_iter=10, heatmap_path='.', max_iters=10, gamma=0.01, fairness_def='FP', predictor=LinearRegression())

Initialize Model Object and set hyperparameters.

Parameters:

• C – Maximum L1 Norm for the Dual Variables (hyperparameter)

• printflag – Print Output Flag

• heatmapflag – Save Heatmaps every heatmap_iter Flag

• heatmap_iter – Save Heatmaps every heatmap_iter

• heatmap_path – Save Heatmaps path

• max_iters – Time Horizon for the fictitious play dynamic.

• gamma – Fairness Approximation Paramater

• fairness_def – Fairness notion, FP, FN, SP.

• errors – see fit()

• fairness_violations – see fit()

• predictor – Hypothesis class for the Learner. Supports LR, SVM, KR, Trees.

fit(dataset, early_termination=True)

Run Fictitious play to compute the approximately fair classifier.

Parameters:

• dataset – dataset object with its own class definition in datasets folder inherits from class StandardDataset.

• early_termination – Terminate Early if Auditor can’t find fairness violation of more than gamma.

Returns:

Self

generate_heatmap(dataset, predictions, vmin=None, vmax=None, cols_index=[0, 1], eta=0.1)

Helper Function to generate the heatmap at the current time.

Parameters:

• iteration – current iteration

• dataset – dataset object with its own class definition in datasets folder inherits from class StandardDataset.

• predictions – predictions of the model self on dataset.

• vmin – see documentation of heatmap.py heat_map function

• vmax – see documentation of heatmap.py heat_map function

pareto(dataset, gamma_list)

Assumes Model has FP specified for metric. Trains for each value of gamma, returns error, FP (via training), and FN (via auditing) values.

Parameters:

• dataset – dataset object with its own class definition in datasets folder inherits from class StandardDataset.

• gamma_list – the list of gamma values to generate the pareto curve

Returns:

list of errors, list of fp violations of those models, list of fn violations of those models

predict(dataset, threshold=0.5)

Return dataset object where labels are the predictions returned by the fitted model.

Parameters:

• dataset – dataset object with its own class definition in datasets folder inherits from class StandardDataset.

• threshold – The positive prediction cutoff for the soft-classifier.

Returns:

dataset_new – modified dataset object where the labels attribute are the predictions returned by the self model

print_outputs(iteration, error, group)

Helper function to print outputs at each iteration of fit.

Parameters:

• iteration – current iter

• error – most recent error

• group – most recent group found by the auditor

save_heatmap(iteration, dataset, predictions, vmin, vmax)

Helper Function to save the heatmap.

Parameters:

• iteration – current iteration

• dataset – dataset object with its own class definition in datasets folder inherits from class StandardDataset.

• predictions – predictions of the model self on dataset.

• vmin – see documentation of heatmap.py heat_map function

• vmax – see documentation of heatmap.py heat_map function

Returns:

(vmin, vmax)