Blinding
Starting…
Starting…
1.
Intentional Unawareness
The omission of one or more of the variables above, especially the sensitive demographic information like Race and Sex, can increase the model's fairness. The idea is that if a model has no access to sensitive variables, then it cannot discriminate against anyone based on those variables. This strategy is called Blinding.
C. GAIN
RACE
EDU.
2.
Procedural Fairness
Fairness can be largely divided into two broad types: procedural and outcome fairness [1][2]. These exist independently: the satisfaction of one does not guarantee that of the other. A decision or event is procedurally fair if the process leading to or involved in it is fair. For example, an unbiased coin toss is considered procedurally fair because the toss itself is impartial, not biased toward any particular side.
On the other hand, we judge a decision's outcome fairness based on its results. For example, a housing approval process that takes affirmative steps to accommodate low-income families may satisfy some conception of outcome fairness if the resulting housing distribution is more equal or equitable.
It is necessary to separate procedural from outcome fairness because they are independent concepts that may sometimes be in conflict. The equitable housing approval process is conceivably fair outcome-wise but not much so procedure-wise. It is aware of and treats applicants differently based on their demographic identity. In this way, it sacrifices procedural fairness for more outcome fairness, specifically housing equitability [3].
Blinded algorithmic models, on the other hand, suffer from the opposite dichotomy in terms of fairness. They exhibit a higher level of procedural fairness by virtue of not explicitly using any sensitive demographic information. Nonetheless, they still fail to guarantee outcome fairness. This is due to a puzzling phenomenon called information leakage.
3.
Information Leakage
Normally, we would expect that if an algorithmic model has no access to sensitive information, then its predictions will be more equal or equitable. This turns out not to be the case. Try it with the model below.
Starting…
Starting…
The positive prediction rates among male and female applicants are roughly the same whether we use Sex or not. In both cases, male applicants receive positive predictions more frequently than female applicants, at approximately the same ratios.
This happens because of an unexpected leakage. When we removed the variable Sex from the model, information on the applicants' sex wasn't fully removed. It still exists within the remaining variables, like Capital Gain and Education. We say that information has "leaked" from Sex into Capital Gain, Education, and Occupation.
C. GAIN
AGE
SEX
RACE
EDU.
OCC.
This leakage happens due to correlational relationships between variables. For example, men have historically had higher incomes than women. A large part of this is due to unequal divisions of labor and existing stigma against women in the workplace. This means that if an applicant has a high income, they are more likely to be male than female. As such, in a dataset, information on an applicant's sex is partially embedded in their income, in this case the Capital Gain variable. The same goes for other variables like Education and Occupation.
Now, there is not a definite relationship between Sex and these other variables. Many women still have higher levels of income than the male median. Many are accomplished academics with high-paying, respectable jobs. The information leakage that we see here is an aggregative phenomenon. It only arises in large datasets where we often lose sight of individual cases and only consider the overall population's statistics. The leakage is of concern in this case because algorithmic models do learn from these population statistics. If there is some bias in the statistics, then the models will also reflect those biases in their predictions [4].
It is difficult, sometimes impossible, to effectively remove information leakages. First, there is no easy way to find out where sensitive information might be hiding among non-sensitive variables, especially in large and complex datasets. Second, even if we could do so, successfully removing all of the sensitive information, the remaining dataset would not be a good representation of the sample population. For example, to remove the correlation between Sex and Education, we would have to modify values in the Education column. Such tampering of the input data risks creating new, unintentional biases, defeating the point of blinding in the first place.
4.
In the End
The existence of information leakages and our inability to completely remove those leakages mean that blinded models are not much more fair outcome-wise than their non-blinded counterparts. Some algorithm designers consider this a fatal flaw. They take equality or equity as an essential requirement for fairness. In the next chapter, we will explore a different strategy - equality of outcome - that satisfies just that.
References
1.
Miller, D., 2021. Justice. The Stanford Encyclopedia of Philosophy (Fall 2021 Edition), Edward N. Zalta (ed.).
2.
Grgić-Hlača, N., Zafar, M. B., Gummadi, K., Weller, A., 2018. The Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18).
3.
Bogen, M., Rieke, A., Ahmed, S., 2020. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. DOI: 10.1145/3351095.3372877
4.
Pedreshi, D., Ruggieri, S., Turini, F., 2008. Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. DOI: 10.1145/1401890.1401959
Data
Dua, D., Graff, C., 2019. UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science.