Fairness is inherently subjective and contextual, no tool or software can fully "de-bias" data
impossibility theorem: no more than one of the three fairness metrics (risk assignments) of demographic parity, predictive parity, and equalized odds can hold at the same time for a well calibrated classifier
Group Fairness Metrics: "groups should receive similar treatments or outcomes" (e.g. males and females)
Demographic Parity: acceptance rates for all demographic groups must be equal
guideline: four-fifths rule-- hires in one group should be no less than 4/5 of hires of another group
benefits: helps to even out imbalances created by historic bias/discrimination
drawbacks: may result in hiring less qualified people if there is more training data for majority demographics because the model may not be able to identify qualifications as well in minority applicants, so the process of hiring may be more random
Equalized Odds: The probability of a qualified applicant being hired must be the same for all groups
benefits: more likely that all or most of hired applicants are qualified
drawbacks: may not help to even out numbers as well and may even reinforce historic bias because minority groups may have lower numbers of qualified applicants overall due to historic lack of opportunities, also makes it more difficult for qualified applicants of all groups to get the job
Individual Fairness Metrics: "similar people should be treated similarly"
Generalized Entropy Index: measures how evenly members of each group are distributed within the applications (0 represents perfect equality, higher numbers represent higher levels of inequality)
benefits: more fine-grained than group metrics because considers individual differences/similarities within groups; accounts for intersectionality; uses multiple variables to determine "similarity" rather than just one overarching group
drawbacks: defining "similarity" is tricky
There often must be trade-offs between group fairness and individual fairness-- increasing one decreases the other
There even will be trade-offs between demographic parity and equalized odds-- although both are group metrics, we cannot have both at once.
There also often must be trade-offs between fairness and accuracy-- increasing fairness often decreases accuracy, but it is important to think about the real impacts that our models are having on people's lives in addition to the accuracy of the predictions
TLDR: Fairness is complex, and there is no approach that will work in all cases. It is a continual process dependent on the stakeholders and applications.
Kypraiou, S. (2021). What is Fairness? . Feminist AI. Retrieved from https://feministai.pubpub.org/pub/what-is-fairness-