Algorithmic Decision Making with Conditional Fairness

ABSTRACTNowadays fairness issues have raised great concerns in
decision-making systems. Various fairness notions have been proposed
tomeasure the degree to which an algorithm is unfair. In
practice,there frequently exist a certain set of variables we term as
fairvariables, which are pre-decision covariates such as users’
choices.The effects of fair variables are irrelevant in assessing the
fair-ness of the decision support algorithm. We thus define
conditionalfairness as a more sound fairness metric by conditioning on
thefairness variables. Given different prior knowledge of fair
vari-ables, we demonstrate that traditional fairness notations, such
asdemographic parity and equalized odds, are special cases of
ourconditional fairness notations. Moreover, we propose a
DerivableConditional Fairness Regularizer (DCFR), which can be
integratedinto any decision-making model, to track the trade-off
betweenprecision and fairness of algorithmic decision making.
Specifically,an adversarial representation based conditional
independence lossis proposed in our DCFR to measure the degree of
unfairness. Withextensive experiments on three real-world datasets, we
demonstratethe advantages of our conditional fairness notation and DCFR.


https://arxiv.org/pdf/2006.10483v1.pdf
https://arxiv.org/pdf/2006.10483v1.pdf
https://arxiv.org/pdf/2006.10483v1.pdf
https://arxiv.org/pdf/2006.10483v1.pdf
https://arxiv.org/pdf/2006.10483v1.pdf
https://arxiv.org/pdf/2006.10483v1.pdf

--
With over 1.2 billion devices now running Windows 10, customer
satisfaction is higher than any previous version of windows.