Avoiding Disparity Amplification under Different Worldviews

Explore a conference talk that delves into the challenge of avoiding disparity amplification in machine learning systems across different worldviews. Examine the research presented by S. Yeom and M. Tschantz at the FAccT 2021 virtual conference, which introduces a novel approach to fairness in AI. Learn about the concept of total variation distance and its application in construct-based fairness and utility. Discover how the researchers conducted empirical tests to validate their theories, with a particular focus on the Wizardwick worldview. Gain insights into the complexities of maintaining fairness in AI systems while accounting for diverse perspectives and societal constructs.

Introduction
Outline
Machine Learning
Worldviews
Total Variation Distance
ConstructBased Fairness and Utility
Empirical Tests
Wizardwick Worldview