AI Fairness 360 documentation¶
Algorithms¶
aif360.algorithms.preprocessing¶
algorithms.preprocessing.DisparateImpactRemover([…]) |
Disparate impact remover is a preprocessing technique that edits feature values increase group fairness while preserving rank-ordering within groups [1]_. |
algorithms.preprocessing.LFR(…[, k, Ax, …]) |
Learning fair representations is a pre-processing technique that finds a latent representation which encodes the data well but obfuscates information about protected attributes [2]_. |
algorithms.preprocessing.OptimPreproc(…[, …]) |
Optimized preprocessing is a preprocessing technique that learns a probabilistic transformation that edits the features and labels in the data with group fairness, individual distortion, and data fidelity constraints and objectives [3]_. |
algorithms.preprocessing.Reweighing(…) |
Reweighing is a preprocessing technique that Weights the examples in each (group, label) combination differently to ensure fairness before classification [4]_. |
aif360.algorithms.inprocessing¶
algorithms.inprocessing.AdversarialDebiasing(…) |
Adversarial debiasing is an in-processing technique that learns a classifier to maximize prediction accuracy and simultaneously reduce an adversary’s ability to determine the protected attribute from the predictions [5]_. |
algorithms.inprocessing.ARTClassifier(…) |
Wraps an instance of an art.classifiers.Classifier to extend Transformer. |
algorithms.inprocessing.GerryFairClassifier([…]) |
Model is an algorithm for learning classifiers that are fair with respect to rich subgroups. |
algorithms.inprocessing.MetaFairClassifier([…]) |
The meta algorithm here takes the fairness metric as part of the input and returns a classifier optimized w.r.t. |
algorithms.inprocessing.PrejudiceRemover([…]) |
Prejudice remover is an in-processing technique that adds a discrimination-aware regularization term to the learning objective [6]_. |
algorithms.inprocessing.ExponentiatedGradientReduction(…) |
Exponentiated gradient reduction for fair classification. |
algorithms.inprocessing.GridSearchReduction(…) |
Grid search reduction for fair classification or regression. |
aif360.algorithms.postprocessing¶
algorithms.postprocessing.CalibratedEqOddsPostprocessing(…) |
Calibrated equalized odds postprocessing is a post-processing technique that optimizes over calibrated classifier score outputs to find probabilities with which to change output labels with an equalized odds objective [7]_. |
algorithms.postprocessing.EqOddsPostprocessing(…) |
Equalized odds postprocessing is a post-processing technique that solves a linear program to find probabilities with which to change output labels to optimize equalized odds [8]_ [9]_. |
algorithms.postprocessing.RejectOptionClassification(…) |
Reject option classification is a postprocessing technique that gives favorable outcomes to unpriviliged groups and unfavorable outcomes to priviliged groups in a confidence band around the decision boundary with the highest uncertainty [10]_. |
aif360.algorithms¶
algorithms.Transformer(**kwargs) |
Abstract base class for transformers. |
Datasets¶
aif360.datasets¶
Base classes¶
datasets.Dataset(**kwargs) |
Abstract base class for datasets. |
datasets.StructuredDataset(df, label_names, …) |
Base class for all structured datasets. |
datasets.BinaryLabelDataset([…]) |
Base class for all structured datasets with binary labels. |
datasets.StandardDataset(df, label_name, …) |
Base class for every BinaryLabelDataset provided out of the box by aif360. |
datasets.RegressionDataset(df, dep_var_name, …) |
Base class for regression datasets. |
Common datasets¶
datasets.AdultDataset([label_name, …]) |
Adult Census Income Dataset. |
datasets.BankDataset([label_name, …]) |
Bank marketing Dataset. |
datasets.CompasDataset([label_name, …]) |
ProPublica COMPAS Dataset. |
datasets.GermanDataset([label_name, …]) |
German credit Dataset. |
datasets.LawSchoolGPADataset([dep_var_name, …]) |
Law School GPA dataset. |
datasets.MEPSDataset19([label_name, …]) |
MEPS Dataset. |
datasets.MEPSDataset20([label_name, …]) |
MEPS Dataset. |
datasets.MEPSDataset21([label_name, …]) |
MEPS Dataset. |
Explainers¶
aif360.explainers¶
explainers.MetricTextExplainer(metric) |
Class for explaining metric values with text. |
explainers.MetricJSONExplainer(metric) |
Class for explaining metric values in JSON format. |
Fairness Metrics¶
aif360.metrics¶
metrics.DatasetMetric(dataset[, …]) |
Class for computing metrics based on one StructuredDataset. |
metrics.BinaryLabelDatasetMetric(dataset[, …]) |
Class for computing metrics based on a single BinaryLabelDataset. |
metrics.ClassificationMetric(dataset, …[, …]) |
Class for computing metrics based on two BinaryLabelDatasets. |
metrics.SampleDistortionMetric(dataset, …) |
Class for computing metrics based on two StructuredDatasets. |
metrics.MDSSClassificationMetric(dataset, …) |
Bias subset scanning is proposed as a technique to identify bias in predictive models using subset scanning [#zhang16]_. |
aif360.metrics.utils¶
This is the helper script for implementing metrics.
metrics.utils.compute_boolean_conditioning_vector(X, …) |
Compute the boolean conditioning vector. |
metrics.utils.compute_num_instances(X, w, …) |
Compute the number of instances, \(n\), conditioned on the protected attribute(s). |
metrics.utils.compute_num_pos_neg(X, y, w, …) |
Compute the number of positives, \(P\), or negatives, \(N\), optionally conditioned on protected attributes. |
metrics.utils.compute_num_TF_PN(X, y_true, …) |
Compute the number of true/false positives/negatives optionally conditioned on protected attributes. |
metrics.utils.compute_num_gen_TF_PN(X, …) |
Compute the number of generalized true/false positives/negatives optionally conditioned on protected attributes. |
metrics.utils.compute_distance(X_orig, …) |
Compute the distance element-wise for two sets of vectors. |
scikit-learn-Compatible API Reference¶
This is the class and function reference for the scikit-learn-compatible
version of the AIF360 API. It is functionally equivalent to the normal API but
it uses scikit-learn paradigms (where possible) and pandas.DataFrame for
datasets. Not all functionality from AIF360 is supported yet. See
Getting Started
for a demo of the capabilities.
Note: This is under active development. Visit our GitHub page if you’d like to contribute!
aif360.sklearn.datasets: Dataset loading functions¶
The dataset format for aif360.sklearn is a pandas.DataFrame with
protected attributes in the index.
Warning
Currently, while all scikit-learn classes will accept DataFrames as inputs,
most classes will return a numpy.ndarray. Therefore, many pre-
processing steps, when placed before an aif360.sklearn step in a
Pipeline, will cause errors.
Utils¶
datasets.NumericConversionWarning |
Warning used if protected attribute or target is unable to be converted automatically to a numeric type. |
datasets.standardize_dataset(df, *, …[, …]) |
Separate data, targets, and possibly sample weights and populate protected attributes as sample properties. |
Loaders¶
datasets.fetch_adult([subset, data_home, …]) |
Load the Adult Census Income Dataset. |
datasets.fetch_german(*[, data_home, cache, …]) |
Load the German Credit Dataset. |
datasets.fetch_bank(*[, data_home, cache, …]) |
Load the Bank Marketing Dataset. |
datasets.fetch_compas([subset, data_home, …]) |
Load the COMPAS Recidivism Risk Scores dataset. |
datasets.fetch_lawschool_gpa([subset, …]) |
Load the Law School GPA dataset |
datasets.fetch_meps(panel, *[, …]) |
Load the Medical Expenditure Panel Survey (MEPS) dataset. |
aif360.sklearn.metrics: Fairness metrics¶
aif360.sklearn implements a number of fairness metrics for group fairness
and individual fairness. For guidance on which metric to use for a given
application, see our
Guidance page.
Meta-metrics¶
metrics.difference(func, y_true[, y_pred, …]) |
Compute the difference between unprivileged and privileged subsets for an arbitrary metric. |
metrics.ratio(func, y_true[, y_pred, …]) |
Compute the ratio between unprivileged and privileged subsets for an arbitrary metric. |
metrics.intersection(func, y_true[, y_pred, …]) |
Compute an arbitrary metric on all intersectional groups of the protected attributes provided. |
metrics.one_vs_rest(func, y_true[, y_pred, …]) |
Compute an arbitrary difference/ratio metric on all intersectional groups of the protected attributes provided in a one-vs-rest manner. |
Scorers¶
metrics.make_scorer(score_func[, is_ratio]) |
Make a scorer from a ‘difference’ or ‘ratio’ metric (e.g. |
Generic metrics¶
metrics.num_samples(y_true[, y_pred, …]) |
Compute the number of samples. |
metrics.num_pos_neg(y_true[, y_pred, …]) |
Compute the number of positive and negative samples. |
metrics.specificity_score(y_true, y_pred, *) |
Compute the specificity or true negative rate. |
metrics.sensitivity_score(y_true, y_pred[, …]) |
Alias of sklearn.metrics.recall_score() for binary classes only. |
metrics.base_rate(y_true[, y_pred, …]) |
Compute the base rate, \(Pr(Y = \text{pos_label}) = \frac{P}{P+N}\). |
metrics.selection_rate(y_true, y_pred, *[, …]) |
Compute the selection rate, \(Pr(\hat{Y} = \text{pos_label}) = \frac{TP + FP}{P + N}\). |
metrics.smoothed_base_rate(y_true[, y_pred, …]) |
Compute the smoothed base rate, \(\frac{P + \alpha}{P + N + |R_Y|\alpha}\). |
metrics.smoothed_selection_rate(y_true, …) |
Compute the smoothed selection rate, \(\frac{TP + FP + \alpha}{P + N + |R_Y|\alpha}\). |
metrics.generalized_fpr(y_true, probas_pred, *) |
Return the ratio of generalized false positives to negative examples in the dataset, \(GFPR = \tfrac{GFP}{N}\). |
metrics.generalized_fnr(y_true, probas_pred, *) |
Return the ratio of generalized false negatives to positive examples in the dataset, \(GFNR = \tfrac{GFN}{P}\). |
Group fairness metrics¶
metrics.statistical_parity_difference(y_true) |
Difference in selection rates. |
metrics.mean_difference(y_true[, y_pred, …]) |
Alias of statistical_parity_difference(). |
metrics.disparate_impact_ratio(y_true[, …]) |
Ratio of selection rates. |
metrics.equal_opportunity_difference(y_true, …) |
A relaxed version of equality of opportunity. |
metrics.average_odds_difference(y_true, …) |
A relaxed version of equality of odds. |
metrics.average_odds_error(y_true, y_pred, *) |
A relaxed version of equality of odds. |
metrics.class_imbalance(y_true[, y_pred, …]) |
Compute the class imbalance, \(\frac{N_u - N_p}{N_u + N_p}\). |
metrics.kl_divergence(y_true[, y_pred, …]) |
Compute the Kullback-Leibler divergence, \(KL(P_p||P_u) = \sum_y P_p(y)\log\left(\frac{P_p(y)}{P_u(y)}\right)\) |
metrics.conditional_demographic_disparity(y_true) |
Conditional demographic disparity, \(CDD = \frac{1}{\sum_i N_i} \sum_i N_i\cdot DD_i\) |
metrics.smoothed_edf(y_true[, y_pred, …]) |
Smoothed empirical differential fairness (EDF). |
metrics.df_bias_amplification(y_true, y_pred, *) |
Differential fairness bias amplification. |
metrics.between_group_generalized_entropy_error(…) |
Compute the between-group generalized entropy. |
metrics.mdss_bias_scan(y_true, probas_pred) |
DEPRECATED: Change to new interface - aif360.sklearn.detectors.mdss_detector.bias_scan by version 0.5.0. |
metrics.mdss_bias_score(y_true, probas_pred) |
Compute the bias score for a prespecified group of records using a given scoring function. |
Individual fairness metrics¶
metrics.generalized_entropy_index(b[, alpha]) |
Generalized entropy index measures inequality over a population. |
metrics.generalized_entropy_error(y_true, y_pred) |
Compute the generalized entropy. |
metrics.theil_index(b) |
The Theil index is the generalized_entropy_index() with \(\alpha = 1\). |
metrics.coefficient_of_variation(b) |
The coefficient of variation is the square root of two times the generalized_entropy_index() with \(\alpha = 2\). |
metrics.consistency_score(X, y[, n_neighbors]) |
Compute the consistency score. |
aif360.sklearn.detectors: Bias detection methods¶
Methods for detecting subsets for which a model or dataset is biased.
Bias scan¶
detectors.bias_scan(X, y_true, y_pred, …) |
Scan to find the highest scoring subset of records. |
aif360.sklearn.preprocessing: Pre-processing algorithms¶
Pre-processing algorithms modify a dataset to be more fair (data in, data out).
Pre-processors¶
preprocessing.FairAdapt(prot_attr, adj_mat) |
Fair Data Adaptation. |
preprocessing.LearnedFairRepresentations([…]) |
Learned Fair Representations. |
preprocessing.Reweighing([prot_attr]) |
Sample reweighing. |
Meta-Estimator¶
preprocessing.ReweighingMeta(estimator[, …]) |
A meta-estimator which wraps a given estimator with a reweighing preprocessing step. |
aif360.sklearn.inprocessing: In-processing algorithms¶
In-processing algorithms train a fair classifier (data in, predictions out).
In-processors¶
inprocessing.AdversarialDebiasing([…]) |
Debiasing with adversarial learning. |
inprocessing.ExponentiatedGradientReduction(…) |
Exponentiated gradient reduction for fair classification. |
inprocessing.GridSearchReduction(prot_attr, …) |
Grid search reduction for fair classification or regression. |
aif360.sklearn.postprocessing: Post-processing algorithms¶
Post-processing algorithms modify predictions to be more fair (predictions in, predictions out).
Post-processors¶
postprocessing.CalibratedEqualizedOdds([…]) |
Calibrated equalized odds post-processor. |
postprocessing.RejectOptionClassifier([…]) |
Reject option based classification (ROC) post-processor. |
Meta-Estimator¶
postprocessing.PostProcessingMeta(estimator, …) |
A meta-estimator which wraps a given estimator with a post-processing step. |
aif360.sklearn.utils: Utility functions¶
Validation¶
utils.check_inputs(X, y[, sample_weight, …]) |
Input validation for debiasing algorithms. |
utils.check_groups(arr, prot_attr[, …]) |
Get groups from the index of arr. |