This paper proposes a new method for uncovering discrimination in decision making systems with continuous value outcomes when lacking the protected attribute values. We demonstrate our method over race discrimination using name and surname proxies. Also, we use a new method for estimating uncertainty in the disparate treatment evaluation to allow for better judgment when using imprecise proxies for the protected attribute. We carry out tests using synthetic data.
Article ID: 2021L19
Publisher: Canadian Artificial Intelligence Association