Learning Fair Classifiers: A Regularization-Inspired Approach

By Yahav Bechavod, Katrina Ligett. 2017

70 citations recorded by the source corpus.

We present a regularization-inspired approach for reducing bias in learned classifiers. In particular, we focus on binary classification tasks over individuals from two populations, where, as our criterion for fairness, we wish to achieve similar false positive rates in both populations, and similar false negative rates in both populations. As a proof of concept, we implement our approach and empirically evaluate its ability to achieve both fairness and accuracy, using the COMPAS scores data for prediction of recidivism.

Subfields: Health Information Management

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Health Information Management

Citations in the corpus

Learning Adversarially Fair and Transferable Representations (2018, methodology): “In two models, we learned the original Y using an MLP (one regularized for fairness (Bechavod & Ligett, 2017), one not; Transfer-Fair and -Unfair, respectively) and trained for the transfer task on its internal representations.”
Does mitigating ML's disparate impact require disparate treatment? (2017, background): “For example, [BL17], proposes another regularization-based scheme with two penalty terms.”
Algorithmic Fairness (2020, background): “These mechanisms involve modifying the machine learning algorithms to account for fairness during the training time [4, 12, 13, 28, 59, 79, 138, 142, 143].”
Active Fairness in Algorithmic Decision Making (2018, background): “Most recent work focuses on meritocratic notions of fairness, or error rate matching [1, 11], such as requiring population subgroups to have equal false positive rates (FPR), equal false negative rates (FNR), or both, i.”
Does mitigating ML's impact disparity require treatment disparity? (2017, methodology): “Many papers in discrimination-aware ML propose to optimize accuracy (or some other risk) subject to constraints on the resulting level of impact parity as assessed by some metric [3, 7, 11–14].”
Learning with Complex Loss Functions and Constraints (2018, background): “A common approach here is to either introduce a regularization term that penalizes violations in constraints, or to formulate an unconstrained weighted version of the problem, and replace the non-continuous portions of the objective and constraints with continuous surrogate functions [15, 3, 33, 4].”
Bias in word embeddings (2020, background): “• What definitions of fairness and discrimination are appropriate and under what conditions? [15, 62] • At which part of an algorithm does bias emerge and in what form? [42, 85, 93] • What are the actual consequences of biased algorithms and who is accountable for them? [6, 68, 76] • How can researc”
Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer (2017, methodology): “11: DIY=i(Y,A, Ŷ ) = |EŶ∼Ber(p)(Ŷ = 1− Y |A = 0, Y = i)− EŶ∼Ber(p)(Ŷ = 1− Y |A = 1, Y = i)| D(Y,A, Ŷ ) = 1 2 (DIY=0(Y,A, Ŷ ) +DIY=1(Y,A, Ŷ )) (12) Our regularization scheme is similar to Bechavod and Ligett [2], Kamishima et al. [25]; see Appendix B for results confirming the efficacy of thi”
Fair Forests: Regularized Tree Induction to Minimize Model Bias (2017, methodology): “Others have taken to creating a differentiable regularization term, and applying it to models such as Logistic and Linear Regression [1, 2, 6, 18].”
Predict Responsibly: Increasing Fairness by Learning To Defer (2017, background): “, 2016) and yielding a regularizer - see Appendix B for results echoing (Kamishima et al., 2012; Bechavod and Ligett, 2017).”