Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Algorithmic Fairness in Performative Policy Learning: Escaping the Impossibility of Group Fairness

Seamus Somerstep, Ya'acov Ritov, Yuekai Sun

TL;DR

Algorithmic fairness practices that leverage performativity to achieve stronger group fairness guarantees in social classification problems (compared to what is achievable in non-performative settings) and leverage the policymaker's ability to steer the population to remedy inequities in the long term are developed.

Abstract

In many prediction problems, the predictive model affects the distribution of the prediction target. This phenomenon is known as performativity and is often caused by the behavior of individuals with vested interests in the outcome of the predictive model. Although performativity is generally problematic because it manifests as distribution shifts, we develop algorithmic fairness practices that leverage performativity to achieve stronger group fairness guarantees in social classification problems (compared to what is achievable in non-performative settings). In particular, we leverage the policymaker's ability to steer the population to remedy inequities in the long term. A crucial benefit of this approach is that it is possible to resolve the incompatibilities between conflicting group fairness definitions.

Algorithmic Fairness in Performative Policy Learning: Escaping the Impossibility of Group Fairness

TL;DR

Algorithmic fairness practices that leverage performativity to achieve stronger group fairness guarantees in social classification problems (compared to what is achievable in non-performative settings) and leverage the policymaker's ability to steer the population to remedy inequities in the long term are developed.

Abstract

In many prediction problems, the predictive model affects the distribution of the prediction target. This phenomenon is known as performativity and is often caused by the behavior of individuals with vested interests in the outcome of the predictive model. Although performativity is generally problematic because it manifests as distribution shifts, we develop algorithmic fairness practices that leverage performativity to achieve stronger group fairness guarantees in social classification problems (compared to what is achievable in non-performative settings). In particular, we leverage the policymaker's ability to steer the population to remedy inequities in the long term. A crucial benefit of this approach is that it is possible to resolve the incompatibilities between conflicting group fairness definitions.
Paper Structure (24 sections, 17 theorems, 55 equations, 2 figures, 1 algorithm)

This paper contains 24 sections, 17 theorems, 55 equations, 2 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Group fairness in performative policy learning
  4. Standard fairness constraints are insufficient in performative prediction
  5. Equality of Outcomes
  6. Feasibility of equality of outcomes
  7. Impossibility Results
  8. Alternative performative models
  9. Feasibility of Equality of Treatment in Alternative Models
  10. A reduction algorithm for equality of outcomes
  11. Algorithm 1 in the Modified Labor Market Model
  12. Summary and discussion
  13. Theoretical Properties of Algorithm \ref{['alg: ex-post-reduction']}
  14. Coate and Loury Result
  15. Section 2 and Section 3 Proofs
  16. ...and 9 more sections

Key Result

Theorem 2.2

It is impossible to find a joint distribution on $(f(X), Y, {G})$ that satisfies two of DP, separation, and sufficiency, unless one of the following hold:

Figures (2)

  • Figure 1: Sufficiency + separation simultaneously
  • Figure 2: Equality of treatment + Equality of responses on training set

Theorems & Definitions (29)

  • Example 2.1: Continuous labor market example somerstep2023Learning
  • Theorem 2.2: chouldechova2017Fairkleinberg2016Inherent
  • Definition 2.3: Equality of outcomes
  • Definition 2.4: equality of responses
  • Definition 2.5: Equality of treatment
  • Proposition 2.6
  • Definition 3.1: stochastic dominance
  • Theorem 3.2
  • Example 3.3: causal strategic classification shravit-caus-strat-LR-2020
  • Example 3.4: Modified Labor Market Model
  • ...and 19 more