Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

State of the Art in Fair ML: From Moral Philosophy and Legislation to Fair Classifiers

Elias Baumann, Josef Lorenz Rumberger

TL;DR

This paper surveys how fairness concerns arise in ML-driven decisions, surveys ethical and legal foundations, and compares observational and causal approaches to detecting and preventing discrimination. It discusses formal and substantive Equality of Opportunity, GDPR and GAET-influenced protections, and multiple fairness criteria such as demographic parity, equalized odds, and calibration. It reviews causal inference methods, counterfactual fairness, and interpretability as mechanisms to ensure accountability. The work highlights practical tradeoffs, limitations, and suggests combining observational, causal, and interpretability tools to reduce discrimination in real-world deployments.

Abstract

Machine learning is becoming an ever present part in our lives as many decisions, e.g. to lend a credit, are no longer made by humans but by machine learning algorithms. However those decisions are often unfair and discriminating individuals belonging to protected groups based on race or gender. With the recent General Data Protection Regulation (GDPR) coming into effect, new awareness has been raised for such issues and with computer scientists having such a large impact on peoples lives it is necessary that actions are taken to discover and prevent discrimination. This work aims to give an introduction into discrimination, legislative foundations to counter it and strategies to detect and prevent machine learning algorithms from showing such behavior.

State of the Art in Fair ML: From Moral Philosophy and Legislation to Fair Classifiers

TL;DR

This paper surveys how fairness concerns arise in ML-driven decisions, surveys ethical and legal foundations, and compares observational and causal approaches to detecting and preventing discrimination. It discusses formal and substantive Equality of Opportunity, GDPR and GAET-influenced protections, and multiple fairness criteria such as demographic parity, equalized odds, and calibration. It reviews causal inference methods, counterfactual fairness, and interpretability as mechanisms to ensure accountability. The work highlights practical tradeoffs, limitations, and suggests combining observational, causal, and interpretability tools to reduce discrimination in real-world deployments.

Abstract

Machine learning is becoming an ever present part in our lives as many decisions, e.g. to lend a credit, are no longer made by humans but by machine learning algorithms. However those decisions are often unfair and discriminating individuals belonging to protected groups based on race or gender. With the recent General Data Protection Regulation (GDPR) coming into effect, new awareness has been raised for such issues and with computer scientists having such a large impact on peoples lives it is necessary that actions are taken to discover and prevent discrimination. This work aims to give an introduction into discrimination, legislative foundations to counter it and strategies to detect and prevent machine learning algorithms from showing such behavior.

Paper Structure

This paper contains 36 sections, 4 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Discrimination - A Short Introduction
  3. Equality of Opportunity (EoO)
  4. Formal Equality of Opportunity:
  5. Substantive Equality of Opportunity:
  6. John Rawls' Theory of Justice as Fairness:
  7. Legal Situation - Germany
  8. General Data Protection Regulation (GDPR)
  9. Discriminatory behavior of algorithm
  10. Where does discriminatory behavior of algorithms come from?
  11. Selection and Confirmation Bias:
  12. Limited Features:
  13. Sample Size Disparity:
  14. Discrimination detection and prevention with the observational approach
  15. Literature regarding the observational approach
  16. ...and 21 more sections

Figures (5)

  • Figure 1: Classifier trained to separate real names (positive examples) from pseudonyms
  • Figure 2: ROC Curve and True Postive Rate to Cost curve with equal odds / equal opportunity selection. (Curves were created for this example)
  • Figure 3: Uncalibrated classifier. Logistic regression on income predictions using UCI Adult Census Dataset Dua:2017
  • Figure 4: Indistiguishable examples for observational criteria
  • Figure 5: Class representations and Saliency maps simonyan2013deep