holisticai.bias.mitigation

The holisticai.bias.mitigation module includes preprocessing, inprocessing and postprocessing bias mitigation algorithms.

Pre-processing

Reweighing	Reweighing preprocessing [1]_ weights the examples in each group-label combination to ensure fairness before classification.
LearningFairRepresentation	Learning fair representations [1]_ finds a latent representation which encodes the data well while obfuscates information about protected attributes [1].
CorrelationRemover	CorrelationRemover applies a linear transformation to the non-sensitive feature columns in order to remove their correlation with the sensitive feature columns while retaining as much information as possible (as measured by the least-squares error).
FairletClusteringPreprocessing	Fairlet decomposition [1]_ is a pre-processing approach that computes fair micro-clusters where fairness is guaranteed.
DisparateImpactRemover	Disparate impact remover [1]_ edits feature values to increase group fairness while preserving rank-ordering within groups.

In-processing

ExponentiatedGradientReduction	Exponentiated gradient reduction
GridSearchReduction	Grid Search Reduction technique can be used for fair classification or fair regression.
MetaFairClassifier	The meta algorithm [1]_ takes the fairness metric as part of the input and returns a classifier optimized w.r.t.
PrejudiceRemover	Prejudice Remover
FairKCenterClustering	Fair K-Center Clustering
FairKMedianClustering	Fair K-Median Clustering
FairletClustering	Fairlet Clustering [1]_ inprocessing bias mitigation works in two steps:
VariationalFairClustering	Variational Fair Clustering [1]_ helps you to find clusters with specified proportions of different demographic groups pertaining to a sensitive attribute of the dataset (group_a and group_b) for any well-known clustering method such as K-means, K-median or Spectral clustering (Normalized cut).
FairScoreClassifier	Fair Score Classifier [1]_ generates a classification model that integrates fairness constraints for multiclass classification.
BlindSpotAwareMF	Blind Spot Aware Matrix Factorization
PopularityPropensityMF	Popularity Propensity Matrix Factorization [1]_ address selection biases in recommender systems by using causal inference techniques to provide unbiased performance estimators and improve prediction accuracy.
FairRec	Fair Recommendation System (FairRec) [1]_, exhibes the desired two-sided fairness by mapping the fair recommendation problem to a fair allocation problem; moreover, it is agnostic to the specifics of the data-driven model (that estimates the product-customer relevance scores) which makes it more scalable and easy to adapt [1]_.
AdversarialDebiasing	Adversarial Debiasing
DebiasingLearningMF	Debiasing Learning Matrix Factorization

Post-processing

CalibratedEqualizedOdds	Calibrated equalized odds postprocessing optimizes over calibrated classifier score outputs to find probabilities with which to change output labels with an equalized odds objective.
EqualizedOdds	Equalized odds postprocessing use linear programming to find the probability with which change favorable labels (y=1) to unfavorable labels (y=0) in the output estimator to optimize equalized odds.
RejectOptionClassification	Reject option classification gives favorable outcomes (y=1) to unpriviliged groups and unfavorable outcomes (y=0) to priviliged groups in a confidence band around the decision boundary with the highest uncertainty.
LPDebiaserBinary	Linear Programmin Debiaser [1]_ is a postprocessing algorithms designed to debias pretrained classifiers.
LPDebiaserMulticlass	Linear Programming Debiaser [1]_ is a postprocessing algorithms designed to debias pretrained classifiers.
MLDebiaser	MLDebiaser postprocessing [1]_ debias predictions w.r.t.
PluginEstimationAndCalibration	Plugin Estimation and Calibration postprocessing [1]_ optimizes over calibrated regressor outputs via a smooth optimization.
WassersteinBarycenter	Fair Regression with Wasserstein Barycenters learning a real-valued function that satisfies the Demographic Parity constraint [1]_ .
DebiasingExposure	Disparate Exposure Learning to Rank (DELTR) [1]_ incorporates a measure of performance and a measure of disparate exposure into its loss function.
FairTopK	Fair Top K bias mitigation [1]_ can be used for Recommender Systems.
MCMF	Minimal Cluster Modification for Fairnes (MCMF) [1]_ is focused on the minimal change it so that the clustering is still of good quality and fairer.
DisparateImpactRemoverRS	Disparate impact remover edits feature values to increase group fairness while preserving rank-ordering within groups.