[Paper Review] Debiasing representations by removing unwanted variation due to protected attributes

Amanda Bower, Laura Niss+2 more|arXiv (Cornell University)|Jul 2, 2018

Explainable Artificial Intelligence (XAI)10 references6 citations

TL;DR

This paper proposes a regression-based preprocessing method to debias machine learning representations by removing unwanted variation due to protected attributes such as race. Using a matrix-variate factor model inspired by the RUV framework in genomics, the approach achieves first-order conditional parity, significantly reducing racial disparities in recidivism risk scores with minimal accuracy loss.

ABSTRACT

We propose a regression-based approach to removing implicit biases in representations. On tasks where the protected attribute is observed, the method is statistically more efficient than known approaches. Further, we show that this approach leads to debiased representations that satisfy a first order approximation of conditional parity. Finally, we demonstrate the efficacy of the proposed approach by reducing racial bias in recidivism risk scores.

Motivation & Objective

• Address algorithmic bias in machine learning representations, particularly in criminal justice risk assessment.
• Reduce disparities in false positive and false negative rates between racial groups in recidivism prediction.
• Develop a preprocessing method that removes unwanted variation due to protected attributes (e.g., race) without requiring retraining of downstream models.
• Ensure debiased representations satisfy a first-order approximation of conditional parity, a fairness criterion.
• Demonstrate the method's efficacy on real-world data, specifically the ProPublica COMPAS dataset.

Proposed method

• Model representations using a matrix-variate factor model where Y = XA^T + ZB^T + E, with X as protected attributes (e.g., race), Z as permissible attributes, and E as error terms.
• Assume a linear relationship between permissible attributes Z and protected attributes X via Z = XΓ^T + W, allowing for confounding between X and Z.
• Estimate the coefficient matrix A using regression of Y on X, assuming control variables (Z) are unobserved but modeled via latent factors.
• Debias representations by regressing out the influence of protected attributes X on Y, conditioned on Z, using a modified RUV approach.
• Apply the debiasing transformation Y_db = Y - XÂ, where  is the estimated coefficient of X in predicting Y, to produce conditionally uncorrelated representations.
• Validate the method using logistic regression on debiased features to predict recidivism, comparing fairness metrics before and after debiasing.

Experimental results

Research questions

• RQ1 Can a regression-based preprocessing method effectively reduce racial bias in recidivism risk scores without sacrificing predictive performance?
• RQ2 Does removing variation due to protected attributes lead to representations that satisfy conditional parity in fairness metrics?
• RQ3 How does the proposed method compare to existing debiasing approaches in terms of statistical efficiency and fairness improvement?
• RQ4 To what extent does the method reduce disparities in false positive and false negative rates between African-American and Caucasian defendants?
• RQ5 Can the method be applied effectively when the permissible attributes (Z) are unobserved, relying only on observed protected attributes (X) and representations (Y)?

Key findings

• The debiased logistic regression model reduced the false negative rate (FNR) disparity between African-Americans and Caucasians from nearly 20% to 4% when using the 80th quantile threshold.
• After debiasing, the false positive rate (FPR) disparity was reduced from 11% to 9% for African-Americans, though Caucasians experienced a slight increase in FPR.
• The model maintained comparable accuracy (65%) on debiased data versus raw data, with a small drop from 67% to 65% at the 50th quantile, indicating minimal performance cost.
• The distribution of recidivism probabilities for African-Americans shifted from being right-skewed in raw representations to more balanced in debiased representations.
• The method achieved a first-order approximation of conditional parity, as the debiased representations were conditionally uncorrelated with protected attributes given the permissible attributes.
• The approach demonstrated statistical efficiency and robustness in reducing bias, particularly in high-stakes decision-making contexts like criminal justice risk assessment.