As machine learning has made its way into more and more areas of our lives, concerns about algorithmic bias have escalated. Machine learning models, which today facilitate decisions about everything from hiring and lending to medical diagnosis and criminal sentencing, may appear to be data-driven and impartial, at least to naïve users—but the typically opaque models are only as good the data they are trained on, and only as ethical as the value judgments embedded in the algorithms.
The burgeoning field of algorithmic fairness, part of the much broader field of responsible computing, is aiming to remedy the situation. For several years now, along with philosophers, legal scholars, and experts in other fields, computer scientists have been tackling the issue. As Stanford University computer science professor Omer Reingold likes to put it, "We are part of the problem, and we should be part of the solution."
Since last year, Reingold has been part of a group of theoretical computer scientists working together through the Simons Collaboration on the Theory of Algorithmic Fairness, funded by the Simons Foundation. Made up of 13 principal investigators, the collaboration is aiming to create a language with which to discuss fairness—"a rigorous language that could translate to actual algorithms," says Reingold, who directs the collaboration.
The Many Meanings of Fairness
The basic building blocks of such a language, just as with a mathematical language for cryptography or privacy, are formal definitions. "When we send a message from one person to another with encryption, we want it to stay secret—but 'secret' is English," Reingold explains. "What cryptography has found is that this could translate to many, many formal definitions that would mean different things."
Although the theory of algorithmic fairness is still in its infancy, with more questions than answers, one thing is abundantly clear already: like the idea of secrecy in cryptography, the English word "fairness" will need a multitude of definitions. No single definition can possibly capture most of what different people mean by that complex and lofty concept. In fact, one collaboration member, Aaron Roth, a computer science professor at the University of Pennsylvania, expresses unease about even using "fairness," a term referring broadly to the distribution of outcomes in the real world, when discussing narrowly scoped measures of disparity within an algorithm. "I think as the field matures, we will start talking less about 'fairness' writ large—a word that it's not clear what it means—and we will start talking more precisely about different kinds of technical disparities that we can talk about methods for eliminating."
Part of what makes fairness such a slippery concept is that it is context-dependent: it can be reasonable to discriminate in one setting and not in another. As collaboration member Cynthia Dwork, Gordon McKay Professor of Computer Science at Harvard University, once put it, "discriminating in advertising for hair products makes perfect sense in a way that discriminating in advertising for financial products is completely illegal."
Also, fairness is typically meaningful only in a relative sense. "Treatment of a particular person or group can sometimes only be judged to be unfair when compared with the treatment of other individuals or groups," says Katrina Ligett, another collaboration member, a professor of computer science at Israel's Hebrew University of Jerusalem, and head of the university's program on the Internet and Society. In these relative notions of fairness, Ligett explains in game-theoretic terms, "The utility of player A depends not just on player A's treatment, but also on B's; and B's utility in turn depends on how player A is treated." That makes fairness different from, say, privacy, which can be rigorously defined in an absolute sense. As Ligett puts it, "It is well-defined to talk about how much privacy A is getting without needing to understand how B's data is being treated."
Clear definitions are not the entirety of a theory of algorithmic fairness—but they are absolutely essential. For one thing, without agreed-upon definitions, people who design software can make vague, baseless claims about their systems. What does it mean to say that a system is secure, or that it is private, or that it is fair? "Without a concrete definition we see that it doesn't mean a lot," Reingold says. Once you have a set of definitions, however, you can prove whether a particular algorithm satisfies a particular definition. Also, you can try to translate each formal definition into plain English so a system's users and policy makers can understand and debate it. These steps may have their own challenges, "But my feeling," Reingold says, "is that if you don't have definitions, you don't have anything."
Like the idea of secrecy in cryptography, the English word "fairness" will need a multitude of definitions.
For these definitions to be useful in the real world, not just mathematically appealing, they must capture important aspects of fairness as conceived of by experts in fields such as ethics, philosophy, and the law. To that end, Ligett has been collaborating with legal scholars. She notes that these conversations are not simple to have. "Your typical law scholar doesn't have her own mathematical definition of fairness that we can line up and compare with a computer scientist's notion."
A theory of algorithmic fairness will also need a taxonomy that is "more than just a giant collection of unrelated definitions," says Roth. The field has the start of such a taxonomy already, as when fairness researchers speak of "families" of definitions. One major family is group fairness, in which statistical measures of the algorithm, such as false positive rates or false negative rates, are equalized across populations. A simple example of group fairness is when the acceptance rate into a university is the same for all races of applicants. Another family, individual fairness, aims for fairness guarantees to apply to individuals rather than to groups, such as when the admissions process treats two similar individuals similarly. Recognizing the strengths and weaknesses of each of those approaches to fairness, several researchers, including Reingold and Roth, came up with a third family of notions, multigroup fairness, that tries to attain the best of both worlds.
A taxonomy helps suggest which kind of definition might be most appropriate for a given application, explains Roth, offering an example from credit-card lending. If a machine-learning system is used to determine whether to increase a borrower's credit limit, instead of just aiding a downstream decision-maker, then it might be sensible to focus on fairness definitions aimed at preventing harms (such as rejecting creditworthy borrowers) from disproportionately falling on one population over another.
The Limits of the Possible
With definitions in place, computer scientists can determine what is possible under given definitions. Knowing that is critical for enabling policy makers to make trade-offs according to the policy makers' values, such as the trade-off between fairness to groups or individuals on the one hand and, on the other, predictive accuracy for the population as a whole.
Emma Pierson, a computer science professor at Jacobs Technion-Cornell Institute at Cornell Tech and Israel's Technion who has studied both the theoretical and applied sides of algorithmic fairness, illustrates this fairness-accuracy trade-off with a stylized example. A healthcare provider might face a choice between two algorithms: one algorithm that is 40% accurate on both Black and White patients (fair in that there's no disparity in accuracy between the two groups) and an algorithm that is 80% accurate on White patients and 60% accurate on Black patients. Which algorithm to choose calls for a value judgment, Pierson says, "But it's clear that if you define fairness as the difference in accuracy across groups, one of these is better at fairness, and one of these is better at overall accuracy." Some algorithms are closer to the Pareto frontier than others—that is, better at both goals—but no algorithm can escape the frontier.
The choice of algorithm also can entail a trade-off among different definitions of fairness, as became clear after computer scientists dug into a dataset tied to the COMPAS algorithm, which courts use to inform decisions about which defendants to release before trial. COMPAS attracted scientists' interest after a 2016 journalistic investigation, published by ProPublica, concluded from its dive into the data that COMPAS was biased against Black defendants; ProPublica's evidence was that in assigning risk scores, COMPAS produced much higher false-positive rates for Blacks than for Whites, meaning that for defendants who did not go on to reoffend, Blacks got classified as likely reoffenders much more often than Whites did. Yet, oddly enough, as computer scientists studying the same data later found, COMPAS satisfied a different definition of fairness: equal calibration for the two populations, meaning that any given COMPAS risk score means the same thing for Black defendants that it does for White defendants—so, for example, a score of 7 translates to the same percentage of defendants reoffending regardless of their race. What's more, the scientists proved that under most real-world conditions, it is impossible to simultaneously satisfy both of these definitions of fairness.
It is important to keep in mind that such impossibility results are not statements about computer algorithms, points out Pierson, who is not part of the Simons collaboration. Certain combinations of desired outcomes are unattainable by anyone who has to make decisions, be they human or machine. "We should be careful: which of the things we're criticizing apply to algorithms specifically, and are uniquely bad with algorithms, as opposed to all decision makers?" Rather than comparing an algorithm to an impossible ideal, we must consider how good the decisions would be if a human were making them, says Pierson. Since humans are notoriously flawed decision-makers, algorithms can often do at least as well.
Of course, they can and should do better. The people who design algorithms can be much more aware of fairness considerations than they are now—and defining what fairness means is a crucial step in that direction. Algorithmic fairness is a messier, more scattered theoretical field than other computer scientists have studied, says Ligett. "But this mess," she adds, "is quite attractive to those who enjoy using the tools of mathematics to try to impose order on things."
Chouldecheva, A. and Roth, A.
A snapshot of the frontiers of fairness in machine learning, Communications, Volume 63, Issue 5, May 2020, pp 82–89 https://doi.org/10.1145/3376898
Dwork, C., Hardt, M., Pitassi, T., Ringold, O., and Zemel, R.
Fairness Through Awareness, ITCS '12: Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, January 2012, Pages 214–226 https://doi.org/10.1145/2090236.2090255
Corbett-Davies, S., Pierson, E., Feller, A., Goel, S., and Huq, A.
Algorithmic decision making and the cost of fairness, KDD '17: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2017, Pages 797–806 https://doi.org/10.1145/3097983.3098095
Kleinberg, J., Mullainathan, S., and Raghavan, M.
Inherent Trade-Offs in the Fair Determination of Risk Scores arXiv:1609.05807 [cs.LG], Thu, 17 Nov 2016 https://arxiv.org/abs/1609.05807v2
Fair prediction with disparate impact: A study of bias in recidivism prediction instruments arXiv:1703.00056, 28 Feb 2017 https://arxiv.org/pdf/1610.07524.pdf
Kearns, M.K., Neel, S., Roth, A., and Wu, Z.S.
Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness, Proceedings of the 35th International Conference on Machine Learning, PMLR 80:2564-2572, 2018 https://proceedings.mlr.press/v80/kearns18a.html
Hébert-Johnson, U., Kim, M., Reingold, O., and Rothblum, G.N.
Multicalibration: Calibration for the (Computationally-Identifiable) Masses, Proceedings of the 35th International Conference on Machine Learning, PMLR 80:1939-1948, 2018 https://proceedings.mlr.press/v80/hebert-johnson18a.html
©2022 ACM 0001-0782/22/8
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or fee. Request permission to publish from [email protected] or fax (212) 869-0481.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2022 ACM, Inc.