Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Efficiency to Equity: Measuring Fairness in Preference Learning

Shreeyash Gowaikar, Hugo Berard, Rashid Mushkani, Shin Koseki

TL;DR

A novel framework for evaluating epistemic fairness in preference learning models inspired by economic theories of inequality and Rawlsian justice is introduced, and metrics adapted from the Gini Coefficient, Atkinson Index, and Kuznets Ratio are proposed to quantify fairness in these models.

Abstract

As AI systems, particularly generative models, increasingly influence decision-making, ensuring that they are able to fairly represent diverse human preferences becomes crucial. This paper introduces a novel framework for evaluating epistemic fairness in preference learning models inspired by economic theories of inequality and Rawlsian justice. We propose metrics adapted from the Gini Coefficient, Atkinson Index, and Kuznets Ratio to quantify fairness in these models. We validate our approach using two datasets: a custom visual preference dataset (AI-EDI-Space) and the Jester Jokes dataset. Our analysis reveals variations in model performance across users, highlighting potential epistemic injustices. We explore pre-processing and in-processing techniques to mitigate these inequalities, demonstrating a complex relationship between model efficiency and fairness. This work contributes to AI ethics by providing a framework for evaluating and improving epistemic fairness in preference learning models, offering insights for developing more inclusive AI systems in contexts where diverse human preferences are crucial.

From Efficiency to Equity: Measuring Fairness in Preference Learning

TL;DR

A novel framework for evaluating epistemic fairness in preference learning models inspired by economic theories of inequality and Rawlsian justice is introduced, and metrics adapted from the Gini Coefficient, Atkinson Index, and Kuznets Ratio are proposed to quantify fairness in these models.

Abstract

As AI systems, particularly generative models, increasingly influence decision-making, ensuring that they are able to fairly represent diverse human preferences becomes crucial. This paper introduces a novel framework for evaluating epistemic fairness in preference learning models inspired by economic theories of inequality and Rawlsian justice. We propose metrics adapted from the Gini Coefficient, Atkinson Index, and Kuznets Ratio to quantify fairness in these models. We validate our approach using two datasets: a custom visual preference dataset (AI-EDI-Space) and the Jester Jokes dataset. Our analysis reveals variations in model performance across users, highlighting potential epistemic injustices. We explore pre-processing and in-processing techniques to mitigate these inequalities, demonstrating a complex relationship between model efficiency and fairness. This work contributes to AI ethics by providing a framework for evaluating and improving epistemic fairness in preference learning models, offering insights for developing more inclusive AI systems in contexts where diverse human preferences are crucial.

Paper Structure

This paper contains 30 sections, 10 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness in Classification and Regression
  4. Fairness in Preference Learning and Ranking
  5. Towards Diverse Preference Representation
  6. Background
  7. Problem Definition
  8. Equality Metrics
  9. Maximal Error Gap
  10. Standard Deviation of the Error
  11. Gini Coefficient gini1936measure
  12. Generalised Entropy Index shorrocks1980class
  13. Atkinson Index atkinson1970measurement
  14. Kutznets Ratio kuznets2019economic
  15. Experimental Setup
  16. ...and 15 more sections

Figures (3)

  • Figure 1: Voting patterns observed in the AI-EDI-Space Dataset. We plot the histogram of the absolute scores given by 3 participants with clearly distinct voting patterns. This figure highlights that users might have very different "taste profiles".
  • Figure 2: Experiment results on the AI-EDI Space Dataset
  • Figure 3: Performance and Metrics for Jester Jokes Dataset over epochs. A clear tradeoff in performance and equality can be seen from the figure where equality seems to increase over training epochs and performance decreases.

Theorems & Definitions (2)

  • Definition 1: Efficiency
  • Definition 2: Equality