From Parity to Preference-based Notions of Fairness in Classification

By Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez-Rodriguez, Krishna P. Gummadi, Adrian Weller. 2017

225 citations recorded by the source corpus.

The adoption of automated, data-driven decision making in an ever expanding range of applications has raised concerns about its potential unfairness towards certain social groups. In this context, a number of recent studies have focused on defining, detecting, and removing unfairness from data-driven decision systems. However, the existing notions of fairness, based on parity (equality) in treatment or outcomes for different social groups, tend to be quite stringent, limiting the overall decision making accuracy. In this paper, we draw inspiration from the fair-division and envy-freeness literature in economics and game theory and propose preference-based notions of fairness -- given the choice between various sets of decision treatments or outcomes, any group of users would collectively prefer its treatment or outcomes, regardless of the (dis)parity as compared to the other groups. Then, we introduce tractable proxies to design margin-based classifiers that satisfy these preference-based notions of fairness. Finally, we experiment with a variety of synthetic and real-world datasets and show that preference-based fairness allows for greater decision accuracy than parity-based fairness.

Subfields: Evaluation & Benchmarks, AI Ethics, Fairness & Society

How researchers see this paper

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

In Evaluation & Benchmarks

In AI Ethics, Fairness & Society

Citations in the corpus

Attack of the Tails: Yes, You Really Can Backdoor Federated Learning (2020, background): “In Figure 7, we report Accuracy Parity Ratio := min pi max pi , where pi is the accuracy on client i’s data [87].”
On formalizing fairness in prediction with machine learning (2017, neutral): “(47) provide following two fairness formalizations.”
Proportionally Fair Clustering (2019, background): “The standard model in fair supervised learning [13, 28, 27, 35, 34] has a set of protected agents given as input to an algorithm which must classify agents into a positive and negative group.”
50 Years of Test (Un)fairness: Lessons for Machine Learning (2018, background): “mise, which in some scenarios might relate to the public’s notion of fairness. We propose that the economics literature on fair division might provide some useful ideas, as has also been suggested by [68]. However, we do heed Darlington’s [19] warning that “a compromise may end up satisfying nobody;”
Aequitas: A Bias and Fairness Audit Toolkit (2018, methodology): “…methods to avoid bias and disparate impact or deﬁning tradeoﬀs among diﬀerent criteria (Calders and Verwer, 2010; Kamishima et al., 2011; Dwork et al., 2012; Zemel et al., 2013; Feldman et al., 2015; Hardt et al., 2016; Kleinberg et al., 2016; Corbett-Davies et al., 2017; Zafar et al., 2017b,a).”
Taking Advantage of Multitask Learning for Fair Classification (2018, background): “In recent years there has been a lot of interest in the problem of enhancing learning methods with “fairness” requirements, see [1– 4 8–10 12 17 19 20 22–26 28 30 31 35–39] and references therein.”
Does mitigating ML's disparate impact require disparate treatment? (2017, background): “The work of [ZVR17] departs from parity-based definitions and proposes instead a preference-based notion of fairness.”
Empirical Risk Minimization under Fairness Constraints (2018, background): “Methods in the second family enforce fairness directly during the training step: [18, 19, 20, 3, 21, 22, 23, 24, 25, 26] are examples of this method which provide non-convex approaches to the solution of the problem or they derive consistency results just for the non-convex formulation resorting…”
Blind Justice: Fairness with Encrypted Sensitive Attributes (2018, methodology): “Zafar et al. (2017c) use a convex approximation of the p % - rule, see eq.”
Explainable AI: A Review of Machine Learning Interpretability Methods (2020, background): “[100], it was pointed out that many of the existing notions of fairness, regarding treatment or impact, are too rigorous and restrictive and, as a result, tend to hinder the overall model performance.”
A Geometric Solution to Fair Representations (2020, background): “In contrast, state-of-the-art fairness methods are specialized to linear regressions or random forests [2, 16, 35].”
Does mitigating ML's impact disparity require treatment disparity? (2017, background): “[30] depart from parity-based definitions and propose instead a preference-based notion of fairness.”
Noise-tolerant fair classification (2019, methodology): “…both on devising sensible measures of fairness (Calders et al., 2009; Dwork et al., 2011; Feldman, 2015; Kim et al., 2018; Hardt et al., 2016; Zafar et al., 2017b,a; Kusner et al., 2017; Speicher et al., 2018; Zhang & Bareinboim, 2018; Heidari et al., 2019), and on means of achieving them…”
Fairness Behind a Veil of Ignorance: A Welfare Analysis for Automated Decision Making (2018, background): “Statistical notions require that given a classifier, a certain fairness metric is equal across all protected groups (see e.g. [Kleinberg et al., 2017; Zafar et al., 2017b,a]).”
Non-Discriminatory Machine Learning through Convex Fairness Criteria (2018, methodology): “It should be noted that (Zafar et al. 2017b) had also compared their method with the non-convex regularization approach proposed in (Kamishima et al. 2012) and had observed a similar trend about the two approaches giving identical results.”
It's COMPASlicated: The Messy Relationship between RAI Datasets and Algorithmic Fairness Benchmarks (2021, background): “5 It is also not uncommon for “granted bail” to be described as the positive decision [105] despite the fact that many individuals are unable to afford bail [106, 107, 108, 109], implying that being incarcerated during the pretrial period is positive for them.”
Fairness in Machine Learning (2020, background): “In many recent papers [1, 3, 6, 14, 15, 16, 24, 44, 46, 52, 69, 90, 91, 92, 96, 97, 128, 149, 153, 177, 183, 184, 186, 188] it has been shown how to enforce EFPRs/EFNRs during model training.”