Philosophy & Technology (2020) 33:225–244 https://doi.org/10.1007/s13347-019-00355-w
RESEARCH ARTICLE
Democratizing Algorithmic Fairness
Pak-Hang Wong1
Copyright By PowCoder代写 加微信 powcoder
Received: 9 September 2018 / Accepted: 14 May 2019 / Published online: 1 June 2019 # Springer Nature B.V. 2019
Machine learning algorithms can now identify patterns and correlations in (big) datasets and predict outcomes based on the identified patterns and correlations. They can then generate decisions in accordance with the outcomes predicted, and decision-making processes can thereby be automated. Algorithms can inherit questionable values from datasets and acquire biases in the course of (machine) learning. While researchers and developers have taken the problem of algorithmic bias seriously, the development of fair algorithms is primarily conceptualized as a technical task. In this paper, I discuss the limitations and risks of this view. Since decisions on Bfairness measure^ and the related techniques for fair algorithms essentially involve choices between competing values, Bfairness^ in algorithmic fairness should be conceptualized first and foremost as a political question and be resolved politically. In short, this paper aims to foreground the political dimension of algorithmic fairness and supplement the current discussion with a deliberative approach to algorithmic fairness based on the accountability for reasonableness framework (AFR).
Keywords Algorithmicbias.Machinelearning.Fairness.Democratization. Accountability for reasonableness
1 Introduction
Friedman and Nissenbaum (1996) have shown that computer systems can be biased, that is, computer systems can Bsystematically and unfairly discriminate against certain individuals or groups of individuals in favor of others^ (Friedman and Nissenbaum 1996, p. 332), and the recognition of bias in computer systems has inspired numerous approaches to detect, scrutinize, and avoid bias in computer systems.1 Despite early
1For an overview of major approaches to assess the values embedded in information technology, see Brey (2010).
* Pak-Hang Wong
Department of Informatics, Universität Hamburg, Vogt-Kölln-Straße 30, 22527 Hamburg,
226 P.-H. Wong
efforts to combat bias in computer systems, the bias in and through computing remains today and possibly in a more problematic form. Machine learning algorithms can now identify patterns and correlations in (big) datasets and predict outcomes based on the identified patterns and correlations. They can then generate decisions in accordance with the outcomes predicted, and decision-making processes can thereby be automated. However, algorithms can inherit questionable values from datasets and acquire biases in the course of (machine) learning (Barocas and Selbst 2016; Mittelstadt et al. 2016). Automated algorithmic decision-making also makes it difficult for people to see algorithms as biased either because they, like big data, invoke Bthe aura of truth, objectivity, and accuracy^ (Boyd and Crawford 2012, p. 663) or because they are incomprehensible to an untrained public, and, worse yet, they can even be inscrutable to trained experts (Burrell 2016; Matthias 2004).
The possible harm from algorithmic bias can be enormous as algorithmic decision- making becomes increasingly common in everyday life for high-stakes decisions, e.g., parole decisions, policing, university admission, hiring, insurance, and credit rating. Several high-profile stories in the media have forcibly directed public attention towards the problem of algorithmic bias, and the public has demanded the industry and research community to create Bfairer^ algorithms.2 In response, researchers and developers have taken the problem seriously and they have proposed numerous methods and techniques to detect and mitigate bias in algorithms (see, e.g., Lepri et al. 2018; Friedler et al. 2019). However, I shall argue that current responses to algorithmic bias are unsatisfac- tory, as the development of fair algorithms is primarily conceptualized as a technical challenge, where researchers and developers attempt to implement some ideas of Bfairness^ within algorithms.
In the next section, I explain in more detail what it is to view algorithmic fairness as a technical challenge and illustrate its limitations and risks. I then elaborate the impossibility theorem about algorithmic fairness and the inherent trade-off between fairness and performance in algorithms and argue that they call for an opening-up of the idea of Bfairness^ in algorithmic fairness, which is political in nature. Since decisions on Bfairness measure^ and the related techniques for fair algorithms essentially involve choices between competing values, Bfairness^ in algorithmic fairness should be con- ceptualized first and foremost as a political question and be resolved politically. I suggest that one promising way forward is through democratic communication. If my characterization of the problem of algorithmic fairness is correct, then the task will not merely be optimizing algorithms to satisfy some fairness measures and improving relevant techniques for fair algorithms but to consider and accommodate diverse, conflicting interests in a society. The aim of this paper, therefore, is to foreground the political dimension of algorithmic fairness and supplement the current discussion with a deliberative approach to algorithmic fairness based on the accountability for reason- ableness framework (AFR) developed by Daniels and Sabin (1997, 2008).
2 The media have reported many cases of (potential) harm from algorithmic decision-making, but the racial bias in the COMPAS recidivism algorithm reported by ProPublica (Angwin et al. 2016; Angwin and Larson 2016), along with Northpointe’s (now renamed to Bequivant^) response to ProPublica’s report (Dieterich et al. 2016), have arguably generated the most discussion. The COMPAS recidivism algorithm has since become the paradigmatic case for research on algorithmic bias, with various research citing it as their motivation or using it as a benchmark. Also, see O’Neil (2016) for an accessible discussion of other cases of algorithmic bias.
Democratizing Algorithmic Fairness 227 2 Algorithmic Fairness Is Not Only a Technical Challenge
A recent survey of measures for measuring fairness and discrimination in algorithms describes the task of algorithmic fairness as Btranslat [ing non-discrimination] regula- tions mathematically into non-discrimination constraints, and develop[ing] predictive modeling algorithms that would be able to take into account those constraints, and at the same time be as accurate as possible^ (Žliobaitė 2017, p. 1061). The quote describes algorithmic fairness primarily as a technical challenge to ensure the outcome of an algorithm to approximate the outcome as required by some fairness criteria, while at the same time maintaining its performance with Bbetter^ programmed algorithms and Bbetter^ pre-processing or post-processing techniques.3 It requires researchers and developers to first presume some ideas of Bfairness^ as a benchmark for their works, e.g., the definition of fairness in non-discrimination regulations, for without accepting some ideas of fairness, it is unclear what researchers and developers are programming into an algorithm and what normative standards they are using to assess whether an algorithm is fair or not.4
From a technical point of view, an agreement on an appropriate understanding of fairness to be programmed into an algorithm is essential to achieve algorithmic fairness. It is essential because an inappropriate (e.g., Bfalse^ or Bincorrect^) understanding of fairness could defeat the result, as an algorithm cannot be fair insofar as the standard it is based on is not fair. Similarly, one can dispute whether an algorithm is fair by questioning the idea of fairness underlying the Bfair^ algorithm in question. For example, the disagreement between ProPublica and Northpointe (now equivant) over whether the COMPAS recidivism algorithm exhibits racial bias can be attributed to their different understandings of fairness or, more precisely, their understandings of a violation of fairness, namely disparate treatment, where protected features are explicitly used in decision-making, and disparate impact, where the result of a decision dispro- portionately impacts the protected groups. Northpointe argues that the algorithm is not biased because the reoffending rate is roughly the same at each COMPAS scale regardless of a defendant’s race; thus, the risk score means the same for different races (Dieterich et al. 2016), whereas ProPublica points out that for those who did not reoffend, black defendants are more likely to be classified as having medium or high risk of reoffending than white defendants. Thus, the algorithm is biased because one group, i.e., black defendants, is systematically subjected to more severe treatment due to the algorithm’s misprediction (Angwin et al. 2016; Angwin and Larson 2016). In this debate, Northpointe and ProPublica have referred to different understandings of the violation of fairness (Corbett-Davies et al. 2016). If there is an agreement on what Bfairness^ stands for, then algorithmic fairness is indeed a technical challenge of finding the best way to program such an idea of Bfairness^ into an algorithm.
3 For a recent overview of the current approaches to algorithmic fairness and different techniques to achieve algorithmic fairness, see Lepri et al. (2018) and Friedler et al. (2019).
4 This is not to claim that the presumed ideas of fairness are unreasonable or idiosyncratic. In fact, some researchers have explicitly referred to the social or legal understandings of fairness in constructing their fairness measures. Still, it is the researchers’ choice to rely on a specific understanding of fairness, but not the others, for their fairness measures, and their choice is rarely informed by the public. I shall return to this point in my discussion of the AFR-based framework.
228 P.-H. , as the dispute on the COMPAS recidivism algorithm demonstrates, the idea of Bfairness^ in algorithmic fairness is far from being uncontested.
The idea of Bfairness^ in algorithmic fairness is in many ways contestable, which present an immediate problem to achieving algorithmic fairness. Firstly, there is a growing number of definitions for what Bfairness^ in algorithmic fairness amounts to, and it seems unlikely for researchers and developers to settle on the definition of fairness anytime soon.5 Secondly, there is a deep disagreement among different philosophical traditions as to what Bfairness^ should mean and what it entails norma- tively (Ryan 2006; Binns 2018a), and the same disagreement exists for the closely related concept of Bequality of opportunity^ as well (Temkin 2017; Arneson 2018). Here, it is useful to reiterate that the disagreement is about the values themselves but not the means to achieve them. Hence, the disagreement cannot be resolved merely by creating Bbetter^ algorithms or using Bbetter^ pre-processing and post-processing techniques, as the normative standard for assessing what counts as Bbetter,^ i.e., the very idea of Bfairness,^ is being the locus of the disagreement. In short, the contestability of Bfairness^ foregrounds the need to settle the meaning of fairness alongside, if not prior to, the technical tasks as described by Žliobaitė.6
What must be emphasized is that the above discussion does not mean to suggest that the industry and research community are unaware of the contestability of fairness and related concepts (see, e.g., Corbett-Davies et al. 2017; Mitchell and Shadlen 2017; Berk et al. 2018; Narayanan 2018). However, there are few attempts to settle the meaning of Bfairness^ and address the questions of what ideas of Bfairness^ are appropriate for algorithmic decision-making and why people should accept them.7 Without addressing such questions, their attempts to create fair(er) algorithms remain at best incomplete.
As the technical challenge to create fair algorithms can only be completed by starting with some understandings of fairness, there is the risk of a closing-down of the critical discussion on the ideas of Bfairness^ in algorithmic fairness, as questioning the meaning of Bfairness^ may hold researchers and developers back from completing the technical tasks by questioning the foundation of their works. In this respect, the focus on technical challenges of algorithmic fairness risks to discourage critical
5 For example, Corbett-Davies et al.’s (2017) analysis of the COMPAS recidivism algorithm refers to three definitions of fairness, i.e., statistical parity, conditional statistical parity, and predictive equality. Berk et al.’s (2018) review of fairness in criminal justice risk assessments refers to six definitions of fairness, i.e., overall accuracy equality, statistical parity, conditional procedure accuracy equality, conditional use accuracy equality, treatment equality, and total fairness. Mitchell and Shadlen’s (2017) recent summary includes 19 definitions of fairness, and a recent talk by (2018) has increased the number of definitions to 21.
6 National or international legislation against discrimination may supply the meaning of fairness to researchers and developers for their design and implementation of algorithms. However, there are two potential short- comings in grounding the Bfairness^ in fair algorithms on national and international legislation. Firstly, the capacity of algorithms to identify patterns and correlations may engender new types of discrimination that are not based on common protected features, e.g., races and genders. Accordingly, the existing legislation is likely to be insufficient. Secondly, national and international legislation is often difficult and slow to change. Therefore, the idea of Bfairness^ in algorithmic fairness is likely to be conservative if it is based on the legislation. Of course, national and international legislation remains important to algorithmic fairness for identifying common types of discrimination.
7 For instance, the reason to opt for a specific definition of fairness is often left unarticulated or implicit in the research, except for a few notable exceptions in which researchers and developers acknowledge or reflect on the normative ground of their choice of definition(s). See, e.g., Dwork et al. (2012) and Lipton et al. (2018).
Democratizing Algorithmic Fairness 229 reflection and opening-up of the definition of fairness for public debate and leads to an
elitist approach to algorithmic fairness (Skirpan and Gorelick 2017).
3 The Impossibility Theorem, the Inherent Trade-off, and the Political
Nature of Algorithmic Fairness
If algorithmic fairness is not merely a technical challenge, then what kind of challenge is algorithmic fairness? Using the impossibility theorem about algorithmic fairness and the inherent trade-off between fairness and performance (or accuracy) in algorithms, I show that algorithmic fairness should not be viewed merely as a technical challenge but also a political challenge.
Recent research has demonstrated that it is mathematically impossible for an algorithm to simultaneously satisfy different popular fairness measures, e.g., disparate treatment and disparate impact, the two fairness measures in the debate on racial bias of the COMPAS recidivism algorithms held by ProPublica and Northpointe, respectively (see, e.g., Friedler et al. 2016; Miconi 2017, Chouldechova 2017; Kleinberg et al. 2017; Berk et al. 2018).8 The impossibility to simultaneously satisfy two (or more) formalized definitions of fairness suggests that no matter how many definitions of fairness we can arrive at, they will remain contestable by some other definitions of fairness. As Friedler et al. point out, the impossibility theorem is Bdiscouraging if one hoped for a universal notion of fairness^ (Friedler et al. 2016, p. 14). The impossibility theorem will also be discouraging to achieving algorithmic fairness from a technical point of view, as no (set of) definition can coherently capture different concerns about fairness at the same time. The immediate lesson from the impossibility theorem is that we need to be more sensitive to the contentious nature of the definition of fairness in the discussion on algorithmic fairness.9
In addition to the impossibility theorem, others have pointed to the inherent trade-off between fairness and performance in algorithms (see, e.g., Corbett-Davies et al. 2017; Berk et al. 2018). The trade-off entails that prioritizing fairness in an algorithm will undermine its performance and vice versa. If the algorithm is intended to promote some social goods, and assuming that when functioning well it can achieve this goal, prioritizing fairness necessarily means a loss in those social goods and thus can be conceived as a cost to the society. For instance, Corbett-Davies et al. (2017) have interpreted the trade-off between fairness and performance in the case of the COMPAS recidivism algorithm as a trade-off between fairness (in terms of disparate impact) and public safety, where optimizing for fairness measures is translated as a failure to detain the medium- to high-risk defendants who are more likely to commit violent crimes and thereby threatening public safety.
8 It is not entirely accurate to describe the incompatibility among different definitions of fairness as Bthe impossibility theorem.^ There are indeed situations where some of the definitions of fairness in question can be satisfied simultaneously, but these situations are highly unrealistic, e.g., when we have perfect predictor or trivial predictor that is either always-positive or always-negative (Miconi 2017).
9 This is not intended to be a knock-down argument against viewing algorithmic fairness primarily as a technical challenge. However, as I have argued the focus on technical tasks can lead to a less critical attitude towards one’s idea of Bfairness,^ it is more likely that researchers and developers who see algorithmic fairness primarily as a technical challenge are less sensitive to the contentious nature of the definition of fairness.
230 P.-H. those who value public safety, fairness measures that significantly reduce public safety will be unacceptable. Moreover, they may argue that fairness measures cannot be genuinely fair when these measures reduce their (public) safety, as optimizing for fairness imposes a risk—or, more precisely, a risk of harm—on people for the benefit of defendants.10 In other words, prioritizing fairness could unfairly put some members of the public at the risk of harm from violent crimes.11 Note that this line of argument can be generalized to other algorithms so long as they are designed and implemented to promote social goods. The inherent trade-off between fairness and performance points to the fact that whether the choice of a fairness measure will be considered as acceptable depends on factors that go beyond the consideration of fairness as narrowly defined in formalized terms, and it will require balancing fairness with other social goods in the process.12
The impossibility theorem and the inherent trade-off between fairness and perfor- mance, therefore, raise the following questions: if researchers and developers cannot simultaneously satisfy two (or more) justified understandings of fairness in an algo- rithm and, at the same time, they have to balance fairness with other social goods, (i) what should they decide on the definition of fairness, the balance between fairness and social goods, etc. for an algorithm? And, more importantly, (ii) how can they justify their decisions to those who will be affected by the algorithm?
To answer these questions, Narayanan (2018) helpfully reminds us that the different fairness measures can be unders
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com