Fairness via Separation of Powers

Abstract

We propose a method of separate actors with mixed incentives for the creation of models which conform to various fairness conditions. Instead of viewing Fair Machine Learning as an optimization problem under fairness constraints, we divide the responsibility into three independently acting groups: a Fairness Condition Creator (Fairness Maximizer), a Model Trainer (Accuracy Maximizer), and an Overseer (Justice Maximizer). By following a workflow inspired by theories of democratic governance and mixed incentive structures, these three groups can converge onto fair and accurate models according to context-specific definitions.