Disparate Impact Remover

Overview

The Disparate Impact Remover is a fairness-aware preprocessing algorithm that aims to remove disparate impact by transforming feature distributions to a common (median) distribution across sensitive attribute groups.

What Problem Does It Solve?

Disparate impact occurs when a seemingly neutral practice has a disproportionate adverse effect on a protected group. The Disparate Impact Remover addresses this issue by modifying the input features in a way that makes it difficult to predict the sensitive attribute from the transformed features, while preserving as much information as possible for the main prediction task.

Key Concepts

• Repair Level: Controls the degree of repair applied to the data (from 0.0 to 1.0):
  • 0.0 = no repair (original data);
  • 1.0 = full repair (maximum fairness, minimum predictability of sensitive attribute);
  • Values between 0 and 1 provide a trade-off between fairness and utility.
• Quantile Transformation: The algorithm works by transforming each feature’s distribution within each sensitive group to match a common median distribution.

How It Works

1. Fit Phase:
   • For each feature, the algorithm computes the empirical cumulative distribution function (CDF) for each sensitive group;
   • It then creates a shared “median” quantile function across groups by taking the median value at each quantile point.
2. Transform Phase:
   • For each feature value in a given sensitive group, the algorithm:
     • Maps the value to its corresponding rank (quantile) in that group’s distribution;
     • Transforms the value by mapping the rank to the corresponding value in the median distribution;
     • Applies the repair level to control how much transformation is applied.
3. Result:
   • The transformed features have distributions that are more similar across sensitive groups;
   • This makes it harder to predict the sensitive attribute from the transformed features.

Implementation Details

The implementation in FairLib uses NumPy for efficient numerical operations and includes the following key components:

• _make_cdf: Creates a function that maps values to their quantiles in a distribution;
• _make_inverse_cdf: Creates a function that maps quantiles to values;
• _transform_feature: Applies the quantile transformation to a single feature.

Usage Example

import numpy as np
from fairlib.preprocessing.disparate_impact_remover import DisparateImpactRemover
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression

# Prepare your data
X_train, X_test, y_train, y_test, s_train, s_test = train_test_split(
    features, labels, sensitive_attributes, test_size=0.2
)

# Initialize the DisparateImpactRemover with desired repair level
dir = DisparateImpactRemover(repair_level=0.8)

# Fit the transformer on training data
dir.fit(X_train, s=s_train)

# Transform both training and test data
X_train_transformed = dir.transform(X_train, s=s_train)
X_test_transformed = dir.transform(X_test, s=s_test)

# Train a classifier on the transformed data
clf = LogisticRegression()
clf.fit(X_train_transformed, y_train)

# Make predictions
y_pred = clf.predict(X_test_transformed)

Advantages and Limitations

Advantages

• Preprocessing approach that works with any downstream classifier;
• Provides a controllable trade-off between fairness and accuracy via the repair level parameter;
• Directly addresses disparate impact in the data.

Limitations

• May reduce overall accuracy of the model;
• Requires sensitive attribute information during both training and inference;
• Works best with numerical features; categorical features need special handling.

References

M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkatasubramanian, “Certifying and removing disparate impact.” ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2015.