Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fairness Through Awareness

Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, Rich Zemel

TL;DR

The paper presents a metric-based framework for fairness in classification, enforcing that similar individuals receive similar outcomes via a Lipschitz constraint and modeling the classifier as a randomized map to outcome distributions. It formulates fairness as a linear program that minimizes expected loss under the Lipschitz constraint, and analyzes connections to statistical parity through Earthmover distance, providing conditions under which individual fairness implies group fairness. It also introduces fair affirmative action via parity-aware optimization, and shows how the approach relates to differential privacy, enabling DP techniques in fairness contexts, with scalable mechanisms under bounded doubling dimension. The discussion covers practical metric design, potential health-care applications (e.g., AALIM-style metrics), and open questions about information leakage and metric specification.

Abstract

We study fairness in classification, where individuals are classified, e.g., admitted to a university, and the goal is to prevent discrimination against individuals based on their membership in some group, while maintaining utility for the classifier (the university). The main conceptual contribution of this paper is a framework for fair classification comprising (1) a (hypothetical) task-specific metric for determining the degree to which individuals are similar with respect to the classification task at hand; (2) an algorithm for maximizing utility subject to the fairness constraint, that similar individuals are treated similarly. We also present an adaptation of our approach to achieve the complementary goal of "fair affirmative action," which guarantees statistical parity (i.e., the demographics of the set of individuals receiving any classification are the same as the demographics of the underlying population), while treating similar individuals as similarly as possible. Finally, we discuss the relationship of fairness to privacy: when fairness implies privacy, and how tools developed in the context of differential privacy may be applied to fairness.

Fairness Through Awareness

TL;DR

The paper presents a metric-based framework for fairness in classification, enforcing that similar individuals receive similar outcomes via a Lipschitz constraint and modeling the classifier as a randomized map to outcome distributions. It formulates fairness as a linear program that minimizes expected loss under the Lipschitz constraint, and analyzes connections to statistical parity through Earthmover distance, providing conditions under which individual fairness implies group fairness. It also introduces fair affirmative action via parity-aware optimization, and shows how the approach relates to differential privacy, enabling DP techniques in fairness contexts, with scalable mechanisms under bounded doubling dimension. The discussion covers practical metric design, potential health-care applications (e.g., AALIM-style metrics), and open questions about information leakage and metric specification.

Abstract

We study fairness in classification, where individuals are classified, e.g., admitted to a university, and the goal is to prevent discrimination against individuals based on their membership in some group, while maintaining utility for the classifier (the university). The main conceptual contribution of this paper is a framework for fair classification comprising (1) a (hypothetical) task-specific metric for determining the degree to which individuals are similar with respect to the classification task at hand; (2) an algorithm for maximizing utility subject to the fairness constraint, that similar individuals are treated similarly. We also present an adaptation of our approach to achieve the complementary goal of "fair affirmative action," which guarantees statistical parity (i.e., the demographics of the set of individuals receiving any classification are the same as the demographics of the underlying population), while treating similar individuals as similarly as possible. Finally, we discuss the relationship of fairness to privacy: when fairness implies privacy, and how tools developed in the context of differential privacy may be applied to fairness.

Paper Structure

This paper contains 35 sections, 14 theorems, 37 equations, 2 figures.

Table of Contents

  1. Introduction
  2. Key Elements of Our Framework
  3. Treating similar individuals similarly.
  4. Formulation as an optimization problem.
  5. Connection between individual fairness and group fairness.
  6. Fair affirmative action.
  7. A close relationship to privacy.
  8. Prevention of certain evils.
  9. Discussion: The Metric
  10. Related Work
  11. Formulation of the Problem
  12. Achieving Fairness
  13. Probability Metrics
  14. Notation.
  15. Useful Facts
  16. ...and 20 more sections

Key Result

Lemma 2.1

Let $D = D_\mathrm{tv}$. Given an instance ${\cal I}$ we can compute $\mathrm{opt}({\cal I})$ with a linear program of size ${\rm poly}(|V|,|A|).$

Figures (2)

  • Figure 1: The Fairness LP: Loss minimization subject to fairness constraint
  • Figure 2: $S_0=G_0\cap S$, $T_0=G_0\cap T$

Theorems & Definitions (35)

  • Definition 2.1: Lipschitz mapping
  • Lemma 2.1
  • Remark 2.1
  • Lemma 2.2
  • proof
  • Lemma 2.3
  • Definition 3.1: Statistical parity
  • Proposition 3.1
  • Definition 3.2: Bias
  • Lemma 3.1
  • ...and 25 more