Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Understanding trade-offs in classifier bias with quality-diversity optimization: an application to talent management

Catalina M Jaramillo, Paul Squires, Julian Togelius

TL;DR

This work addresses how dataset bias shapes classifier fairness and accuracy in talent-management contexts. It introduces a quality-diversity approach based on Covariance Matrix Adaptation MAP-Elites (CMA-ME) to generate and visualize a spectrum of models that vary in bias descriptors, enabling explicit analysis of the fairness–accuracy trade-off. Using Promotion and Adult datasets, the authors demonstrate that unbiased data yield broad regions of high accuracy, while biased data concentrate high accuracy near the dataset's biases; they also show how the best fair model can achieve regulatory fairness (the 80% rule) with modest accuracy losses. The method offers a practical tool for selecting models under fairness constraints and provides a scalable path to more sophisticated bias analyses in real-world decision-making systems.

Abstract

Fairness,the impartial treatment towards individuals or groups regardless of their inherent or acquired characteristics [20], is a critical challenge for the successful implementation of Artificial Intelligence (AI) in multiple fields like finances, human capital, and housing. A major struggle for the development of fair AI models lies in the bias implicit in the data available to train such models. Filtering or sampling the dataset before training can help ameliorate model bias but can also reduce model performance and the bias impact can be opaque. In this paper, we propose a method for visualizing the biases inherent in a dataset and understanding the potential trade-offs between fairness and accuracy. Our method builds on quality-diversity optimization, in particular Covariance Matrix Adaptation Multi-dimensional Archive of Phenotypic Elites (MAP-Elites). Our method provides a visual representation of bias in models, allows users to identify models within a minimal threshold of fairness, and determines the trade-off between fairness and accuracy.

Understanding trade-offs in classifier bias with quality-diversity optimization: an application to talent management

TL;DR

This work addresses how dataset bias shapes classifier fairness and accuracy in talent-management contexts. It introduces a quality-diversity approach based on Covariance Matrix Adaptation MAP-Elites (CMA-ME) to generate and visualize a spectrum of models that vary in bias descriptors, enabling explicit analysis of the fairness–accuracy trade-off. Using Promotion and Adult datasets, the authors demonstrate that unbiased data yield broad regions of high accuracy, while biased data concentrate high accuracy near the dataset's biases; they also show how the best fair model can achieve regulatory fairness (the 80% rule) with modest accuracy losses. The method offers a practical tool for selecting models under fairness constraints and provides a scalable path to more sophisticated bias analyses in real-world decision-making systems.

Abstract

Fairness,the impartial treatment towards individuals or groups regardless of their inherent or acquired characteristics [20], is a critical challenge for the successful implementation of Artificial Intelligence (AI) in multiple fields like finances, human capital, and housing. A major struggle for the development of fair AI models lies in the bias implicit in the data available to train such models. Filtering or sampling the dataset before training can help ameliorate model bias but can also reduce model performance and the bias impact can be opaque. In this paper, we propose a method for visualizing the biases inherent in a dataset and understanding the potential trade-offs between fairness and accuracy. Our method builds on quality-diversity optimization, in particular Covariance Matrix Adaptation Multi-dimensional Archive of Phenotypic Elites (MAP-Elites). Our method provides a visual representation of bias in models, allows users to identify models within a minimal threshold of fairness, and determines the trade-off between fairness and accuracy.

Paper Structure

This paper contains 16 sections, 7 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Bias in talent management and the 80% rule
  4. Methods for handling bias in machine learning
  5. Quality-diversity search
  6. Methodology
  7. Data
  8. Results
  9. Unbiased sample (Figure \ref{['unbiasfig']}, Table \ref{['unbiased results']}).
  10. Male biased sample (Figure \ref{['malebiasfig']}, Table \ref{['male biased results']})
  11. Male biased sample, male oversampling (Figure \ref{['malebiasoverfig']}, Table \ref{['male biased over results']})
  12. Cross biased sample, male young and female old (Figure \ref{['crossbiasfig']},Table \ref{['cross biased results']}).
  13. Adult dataset, unbiased sample (Figure \ref{['adultunbiased']},Table \ref{['adult unbiased results']}).
  14. Adult dataset, stratified sample (Figure \ref{['aacrossbiasfig']},Table \ref{['adult stratified results']}).
  15. Discussion
  16. ...and 1 more sections

Figures (7)

  • Figure 1: Fair zone within the map
  • Figure 2: Unbiased sample map
  • Figure 3: Male biased sample map
  • Figure 4: Male biased and oversampled sample map
  • Figure 5: Cross biased sample map
  • ...and 2 more figures