A Theory of FairnessChris Lam
April 6, 2023 |
AbstractThe fairness problem in machine learning is a sociotechnical problem that requires bringing together central ideas from the social sciences and humanities into mathematics and computer science. In our research, we show how to directly map a principal cause of algorithmic bias (the structure and agency debate in sociology) to Judea Pearl's two fundamental laws of causal inference (counterfactuals / interventions and conditional independence). We represent the supervised machine learning problem as a causal Bayesian network, allowing us to visualize different forms of fairness and discrimination. The goal of this non-partisan framework is to help policymakers on both sides of the aisle to modernize AI regulations so that they can be aligned to society's values.
|