Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Procedural Fairness Through Decoupling Objectionable Data Generating Components

Zeyu Tang, Jialu Wang, Yang Liu, Peter Spirtes, Kun Zhang

TL;DR

The paper tackles disguised procedural unfairness by shifting fairness attention from outcomes to the data generating process itself. It introduces a Rawls-inspired framework that decouples objectionable components from neutral ones using a value instantiation rule and reference-point configurations, ensuring predictions rely only on neutral information where possible. The approach is implemented via modular local causal mechanisms and an ex post optimization to maximize the least advantaged group’s expected outcome, with experiments on simulated data and real datasets (UCI Adult, Folktables) showing improved fairness for disadvantaged groups compared to outcome-centric methods. This modular, transparent framework provides procedural guarantees on fairness of the generation process and offers a scalable path for applying fairness principles to complex causal models in decision-making tasks.

Abstract

We reveal and address the frequently overlooked yet important issue of disguised procedural unfairness, namely, the potentially inadvertent alterations on the behavior of neutral (i.e., not problematic) aspects of data generating process, and/or the lack of procedural assurance of the greatest benefit of the least advantaged individuals. Inspired by John Rawls's advocacy for pure procedural justice, we view automated decision-making as a microcosm of social institutions, and consider how the data generating process itself can satisfy the requirements of procedural fairness. We propose a framework that decouples the objectionable data generating components from the neutral ones by utilizing reference points and the associated value instantiation rule. Our findings highlight the necessity of preventing disguised procedural unfairness, drawing attention not only to the objectionable data generating components that we aim to mitigate, but also more importantly, to the neutral components that we intend to keep unaffected.

Procedural Fairness Through Decoupling Objectionable Data Generating Components

TL;DR

The paper tackles disguised procedural unfairness by shifting fairness attention from outcomes to the data generating process itself. It introduces a Rawls-inspired framework that decouples objectionable components from neutral ones using a value instantiation rule and reference-point configurations, ensuring predictions rely only on neutral information where possible. The approach is implemented via modular local causal mechanisms and an ex post optimization to maximize the least advantaged group’s expected outcome, with experiments on simulated data and real datasets (UCI Adult, Folktables) showing improved fairness for disadvantaged groups compared to outcome-centric methods. This modular, transparent framework provides procedural guarantees on fairness of the generation process and offers a scalable path for applying fairness principles to complex causal models in decision-making tasks.

Abstract

We reveal and address the frequently overlooked yet important issue of disguised procedural unfairness, namely, the potentially inadvertent alterations on the behavior of neutral (i.e., not problematic) aspects of data generating process, and/or the lack of procedural assurance of the greatest benefit of the least advantaged individuals. Inspired by John Rawls's advocacy for pure procedural justice, we view automated decision-making as a microcosm of social institutions, and consider how the data generating process itself can satisfy the requirements of procedural fairness. We propose a framework that decouples the objectionable data generating components from the neutral ones by utilizing reference points and the associated value instantiation rule. Our findings highlight the necessity of preventing disguised procedural unfairness, drawing attention not only to the objectionable data generating components that we aim to mitigate, but also more importantly, to the neutral components that we intend to keep unaffected.
Paper Structure (42 sections, 10 equations, 5 figures, 4 tables, 2 algorithms)

This paper contains 42 sections, 10 equations, 5 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Notations
  4. Requirements for Procedural Fairness
  5. Illustrating Disguised Procedural Unfairness
  6. Decoupling Objectionable Components for Procedural Fairness
  7. Reflecting on a Simple Approach: An Initial Attempt
  8. Our Framework
  9. The Value Instantiation Rule for Local Causal Modules
  10. The Configuration of Reference Point Values
  11. Experiments
  12. Illustrating \ref{['main:rawls:requirement_difference']} Violation: Revisiting Linear Example
  13. Results on UCI Adult Data Set
  14. Concluding Remarks
  15. Previous Works on Algorithmic Fairness Related to Data Generating Process
  16. ...and 27 more sections

Figures (5)

  • Figure 1: The linear example where causal fairness notions are applied. Panel (a) contains the causal graph for the data generating process. Panel (b) and panel (c) summarize the behavior of fitted parameters, where panel (b) corresponds to fairness constraints proposed by kilbertus2017avoiding, and panel (c) corresponds to those proposed by nabi2018fairnabi2019learningnabi2022optimal. Orange solid-line boxes in the matrix are instantiations of the disguised procedural unfairness due to the violation of \ref{['main:rawls:requirement_fair_equality']}.
  • Figure 2: Experimental results on the simulated data and the real-world UCI Adult data set. Panel (a) demonstrates the disguised procedural unfairness due to the violation of \ref{['main:rawls:requirement_difference']}. Panel (b) presents the causal graph for UCI Adult data set. Panel (c) summarizes results on UCI Adult data set, where we present comparisons between group-wise approval rates for low-/high- income individuals, before and after fairness considerations are implemented.
  • Figure 3: Detailed illustrations of the application of value instantiation rule. Panel (a) presents the causal graph, with red edges denoting objectionable components. Panels (b) to (d) present local causal modules that involve objectionable component(s).
  • Figure 4: Causal graph for experiments with simulated and real-world data sets.
  • Figure 5: Different objectionable component configurations in experiments on the UCI Adult data set. Red edges denote potential locations of objectionable components, among which solid-line edges (highlighted with contours) represent objectionable components that we decouple through reference points, and dashed-line edges represent potentially neutral (not objectionable) components.

Theorems & Definitions (5)

  • Definition 4.1: Reference Point
  • Remark C.1
  • Remark C.2
  • Remark C.3
  • Remark C.4