Increasing adoption of machine learning approaches in highstakes domains, such as healthcare and social assistance, has led to increased scrutiny of their impact on vulnerable groups. A number of studies demonstrating the disparate impact of automation on such groups (Citron and Pasquale 2014;Angwin et al. 2016;Dastin 2018;Lee 2018;Koenecke et al. 2020) has motivated an extensive literature that aims at achieving group fairness of machine learning (Kearns et al. 2018;Agarwal et al. 2018;Pleiss et al. 2017;Hardt, Price, and Srebro 2016;Chouldechova and Roth 2018;Mehrabi et al. 2021;Barocas, Hardt, and Narayanan 2017;Angwin et al. 2016;Dwork et al. 2012) by imposing an explicit constraint that prediction efficacy (which can be measured in many different ways) is similar across groups. However, a principal contemplating a change from a conventional, and potentially biased, prediction model to a group-fair approach must contend with the perception that such a switch could inadvertently harm many individuals in the process of improving fairness (for example, making them less likely to receive a scarce resource such as a welfare benefit or admission to a college). Such perceptions could make any change from the status quo contentious and, consequently, less likely. Our central question is whether it is possible to make group-fair classifiers sufficiently popular-reducing the prevalence of realized or perceived harm-so as to make their adoption less contested.
We model the principal's problem as a comparison between a conventional approach f C and a group-fair approach f F , with the principal considering a switch from the former to the latter. Both algorithms select a subset of individuals from a target population to obtain a particular desirable outcome (e.g., a resource, such as admission to a college). We examine popularity in this context through the lens of preferences of individuals in a target population over selection outcomes (which we can encode as positive outcomes of binary classification): an individual weakly prefers f F to f C if the probability of being selected is not lower under the former than under the latter. Popularity of a group-fair approach f F then amounts to ensuring that a given fraction (e.g., majority) of a target population prefers f F to f C .
To illustrate the relationship between fairness, accuracy, and popularity, consider the following example. Let G 1 and G 0 have four and two members respectively, with true labels h1, 1, 1, 1i and h0, 0i. A randomized conventional classifier f C , predicts each member of G 1 to be positive with probability 0.75 and each member of G 0 to be positive with probability 0.25. Under demographic parity fairness, G 1 is advantaged as this group has a positive rate 0.5 greater than that of G 0 . Consider two choices for a fair model. f F1 predicts members of G 1 to be positive with probability 0.75 and members of G 0 to be positive with probability 0.55. f F2 predicts one member of G 1 to be positive with probability 1, and the others with probability 2 3 ; it predicts one member of G 0 to be positive with probability 1 and the other with probability 0.1. Note that both models have identical accuracy and unfairness, namely .65 and .2 respectively. However, f F1 has not decreased the score of any agent in the population; all six prefer f F1 at least as much as the original f C . In contrast, f F2 has decreased the scores of three agents from G 1 and one agent from G 0 ; only two agents prefer f F2 at least as much as f C . This example illustrates that popularity should be viewed as a different axis than either accuracy or fairness, and there may be space to innovate by enabling popularity comparisons among fair(er) models.
We start this paper by asking an empirical question: Do typical group-fair classification approaches yield models that are, in fact, unpopular in the sense above? We demonstrate that they are: in experiments on several standard datasets, more than half the target population can strictly prefer the conventional scheme to several prominent groupfair learning methods. Given that the group-fair approaches have significant motivation and momentum behind them, instead of designing an entirely new approach to finding popular and fair classifier, we ask whether it is possible to minimally postprocess the output of a group-fair classifier in order to achieve some target popularity while maintaining a high level of fairness. We answer this question in the affirmative. Specifically, we describe two approaches to efficiently postprocess the outputs from a given group-fair classifier in order to boost its popularity. The first approach formalizes the problem as a minimal change of outcome probabilities over the target population to guarantee a target level of fairness and popularity. We show that this problem can be solved in polynomial time. Our second approach involves a form of regularized empirical risk minimization with fairness and popularity constraints. This approach relies on partitioning prediction scores into a set of quantiles, and we show that, in general, the problem is strongly NP-Hard. However, we also show that if the number of quantiles is constant, this problem can be solved in polynomial time. Our methods are applicable in both the classification and scarce resource allocation settings, and allow a model designer to directly control the level of popularity and fairness.
In summary, our contributions are: 1. We propose the notion of popularity of group-fair classifiers and allocation schemes, measuring the fraction of a population that is weakly better off when switching from a conventional to a fair learning scheme. 2. We demonstrate the degree to which state of the art group-fair approaches are unpopular compared to their conventional counterparts. 3. We introduce two postprocessing algorithms which allow a principal to directly control the popularity of a given fair model, while maintaining good fairness properties. The first post-processing technique, dubbed DOS (Direct Outcome Shift), is polynomial time solvable for both deterministic and randomized classifiers, and can also be applied to the scarce resource allocation setting. The second technique, k-QLS (k-Quantile Lottery Shift), works by grouping agents into k quantiles (where k is chosen by the model designer), and running lotteries on each quantile. k-QLS is polynomial time solvable for deterministic classifiers. While we show that k-QLS is NP-hard in the randomized case, it becomes tractable for constant k, as would be standard in practice. 4. We empirically demonstrate that the proposed postprocessing techniques can achieve high levels of popularity and fairness with minimal impact on prediction accuracy. Related Work: Our work is broadly related to the field of algorithmic group fairness which is concerned with both defining what it means for a model to be fair, as well as operationalizing these definitions to produce fair models (Hardt, Price, and Srebro 2016;Pleiss et al. 2017;Feldman et al. 2015;Dwork et al. 2012;Agarwal et al. 2018;Kearns et al. 2018;Jang, Shi, and Wang 2021;Kusner et al. 2017). In particular our algorithms work through postprocessing, a common technique for for achieving fairness (Pleiss et al. 2017;Hardt, Price, and Srebro 2016;Kamiran, Karim, and Zhang 2012;Canetti et al. 2019;Lohia et al. 2019;Jang, Shi, and Wang 2021). In these works, the scores or decisions of a conventional classifier are modified in order to achieve fairness. Most post processing techniques for fairness work through "inclusion/exclusion" systems where a potentially randomized procedure is uniformly applied across groups, e.g. random selection of group-specific thresholds (Hardt, Price, and Srebro 2016;Jang, Shi, and Wang 2021), or randomly selecting agents from one group to receive positive classification with constant probability (Pleiss et al. 2017). Our postprocessing techniques, while concerned not exclusively with fairness, follow a similar inclusion/exclusion system.
More generally, randomized prediction methods are common in prior literature. In some cases, randomization is inherently desirable, for example, to explore or correct existing bias in domains such as hiring (Berger et al. 2020;Tassier and Menczer 2008;Hong and Page 2004) or lending (Karlan and Zinman 2010a,b). In other settings, the aim is to increase model robustness (Pinot et al. 2019;Salman et al. 2019), or to achieve better trade-offs between model performance and fairness, as is common in many group-fair classification approaches (Agarwal et al. 2018;Kearns et al. 2018;Pleiss et al. 2017).
Several recent papers look at the potential negative consequences of applying group fairness (Liu et al. 2018;Zhang et al. 2020;Corbett-Davies and Goel 2018;Kasy and Abebe 2021;Ben-Porat, Sandomirskiy, and Tennenholtz 2019). In particular (Liu et al. 2018;Ben-Porat, Sandomirskiy, and Tennenholtz 2019) demonstrate that specific types of group equity can be decreased by the use of fair algorithms. Others have merged notions of welfare and fairness (Hu and Chen 2020;Cousins 2021;Chen and Hooker 2021). Both the notion of popularity, as well as our proposed techniques for satisfying popularity and fairness, differ from these lines of work in that popularity casts welfare in terms of the fraction of a population which prefers a fair model compared with a fairness-agnostic model. While the idea of agent preference over models has received some recent attention (Ustun, Liu, and Parkes 2019) (which aims to classify a population using multiple models such that each agent prefers their assigned model over all others), popularity in the context of group fair learning has remained unexplored thus far.
We begin by formalizing our models of conventional and fair learning, as well as our definition of popularity. Let D =( X,Y,G) be a dataset of n examples where the i th example (x i ,y i ,g i ) consists of features x i 2 X ⇢ R d , binary labels y i 2{ 0, 1} and binary group membership g i 2{ 0, 1}. We assume throughout that the positive label y =1corresponds to the preferred outcome, such as being selected to receive a valuable resource (e.g., college admission). Consider two learning schemes, say C and F , where C is a conventional learning scheme, designed to minimize some fairness-agnostic objective, and F is a fair learning scheme, designed to achieve a desired level of fairness between groups. Then, C solves a problem of the form:
i.e., choosing an optimal model f C from the hypothesis class H C with respect to the loss L C . However, we do not require that f C is the result of strict error minimization, only that it maps X to {0, 1}. In the context of conventional learning, the objective L C and learning scheme C may have exogenous considerations aside from error minimization, such as robustness or interoperability.
We further assume that the learned classifier is of the type that produces a score function h : X ! [0, 1] which is used to induce the classification f (X). Most classifiers used in practice yield such score functions (e.g., SVM, Logistic Regression, Neural Nets, Decision Trees, etc.). We study both deterministic and randomized classifiers in this framework. While deterministic predictions are most common, randomization can offer flexibility that can play a useful role both in achieving fairness (Dwork et al. 2012;Kearns et al. 2018) and robustness (Pinot et al. 2019;Li and Vorobeychik 2015;Salman et al. 2019;Vorobeychik and Li 2014). A deterministic classifier f can be thought of as a threshold on scores from h, i.e., f (x)=I ⇥ h(x) ✓ ⇤ for threshold ✓. A randomized classifier f , in turn, can be viewed as a Bernoulli random variable with a mean given by h, i.e., E[f (x)] = h(x).
In addition to the classification setting above in which, in principle, anyone can be selected (i.e., assigned a positive outcome y =1), we consider scarce resource allocation (henceforth simply allocation). In the allocation setting, unlike the classification setting, the model designer is limited in the number of positive predictions-that is, the number of individuals that can be selected. Specifically, the score function h is used to allocated k<nhomogeneous, indivisible, goods among a population of n agents. This follows a well-established paradigm of allocating scarce resources among individuals using a score function learned on a binary prediction task (Kube, Das, and Fowler 2019). Let I i (X,h,k) 2{ 0, 1} indicate allocation of a resource to agent i when score function h is applied to a population X and there are k resources. Similar to the classification setting, the allocation function I can be deterministic or randomized. In the case of deterministic allocation, I i is obtained directly from the set of scores h(X), e.g., allocating resources to the k highest scoring individuals. In randomized allocation, I i is a Bernoulli random variable, but unlike in the classification setting, I i may have an arbitrary joint relationship with allocation decisions made for other agents, e.g., sampling without replacement weighted by h(X).
Let M(f (X),Y; g) be an efficacy metric computed with respect to group membership g 2{ 0, 1} (for example, false positive rate (FPR) or error rate (ERR)). Define group disparity U (f, D)=|M(f (X),Y; 1) M(f (X),Y; 0)|, i.e., the difference in efficacy between two groups. Then the group-fair learning scheme F solves a problem of the form
i.e., f F is an optimal group-fair model from hypothesis class H F , with fairness captured by the constraint that group disparity U is bounded by . We refer to the fair learning scheme and model f F as -fair. Note that when resources are scarce, fairness is defined over allocation decisions I i (X,h,k), not over scores h; an example of a fairness objective would be selection rate parity of I i (X,h,k) between groups. Our analysis that follows applies to the broad class of additive efficacy metrics in both the classification and allocation settings. Definition 2.1. (Additive Efficacy Metric): An efficacy metric M is additive if for any population (X,Y,G),
for some c
In an additive efficacy metric, the coefficients c
give the respective "costs" of classifying an example from group G g , with true label y, as negative or positive, respectively. Thus, unfairness U is given as the difference in the total efficacy cost between groups. Additive metrics are widely studied in the literature and include metrics such as error rate (ER), positive (or selection) rate (PR), false positive rate (FPR), and true positive rate (TPR). As an example, in the case of PR fairness c We consider the situation in which a conventional learning scheme C is initially in place, and a principal considers a switch from C to a group-fair scheme F , and wishes to ensure that F is -popular in the sense that it is preferred to C by at least a fraction of the target population. We formalize preference over learning schemes by assuming that an individual prefers schemes which yield higher expected outcomes for them, that is, they prefer being selected to not being selected, as in Hardt, Price, and Srebro (2016). Thus, an individual i with features
when decisions are deterministic and
when decisions are stochastic. Note that our analysis is in the space of outcomes, rather than scores. Consequently, if decisions are deterministic, either in classification or allocation settings, agents only have a definitive preference over scores produced by h if this is consequential to outcomes (e.g., pushing them above or below ✓). In the stochastic case, on the other hand, agents prefer the classifier or allocation scheme which yields the higher expected outcome (that is, higher probability of being selected).
Armed with this model of individual preference, we now define what it means for F to be popular.
Definition 2.2. (γ-popularity): A learning scheme F is said to be -popular with respect to a population (X,Y,G) and conventional scheme C, if Condition (3) (for deterministic models), or Condition (4) (for randomized models), holds for at least |X| individuals.
Popularity thus captures the fraction of a population which is weakly better off (or, equivalently, not made worse) from the use of F over C. Similar to the concept offairness, in which a model designer can specify the desired level of fairness , the definition of popularity, as well as our postprocessing techniques described later, allow the model designer to directly specify, and control, the desired level of popularity. Note that we do not capture the degree to which individuals are made better or worse off as a result of switching from C to F , but only whether they are.
As mentioned earlier, our setting is one of a concrete choice by a principal between a particular conventional approach C and a particular group-fair approach F . This reflects a decision by the principal to switch from C-which is currently deployed-to F in order to reduce impact to a disadvantaged group (or groups). Of course, different pairs of C and F (e.g., using different loss functions, different learning algorithms, etc) would yield different judgments about popularity of F , which is, by construction, relative to C. Consequently, these will also yield different decisions about improving popularity of F based on algorithms we discuss below. Nevertheless, our framework generalizes immediately to a setting in which neither C nor F are fixed, and there is uncertain about either, or both. In such a case, we treat uncertainty about either C or F as a distribution over approaches and, consequently, over outcomes induced. This can then be immediately captured within our framework dealing with randomized schemes, and all definitions above, and technical results below, go through unchanged.
Our goal is to investigate the following three questions: 1) Are common group-fair learning techniques popular? 2) For a given and , can we compute -fair and -popular decisions in polynomial time? 3) What is the nature of the tradeoff between popularity, fairness, and accuracy?
We consider two approaches to minimally postprocess afair scheme f F such that the resulting decisions also become -popular, for exogenously specified and : 1) direct outcome shift (DOS) and 2) k-quantile lottery shift (k-QLS). Postprocessing is performed in a transductive setting, in which the populations' features (X,G) (and possibly also labels Y ) are known in advance. Throughout, we use f P to refer to either approach we propose that combines both popularity and group fairness. Direct Outcome Shift (DOS) DOS-based postprocessing arises from solving the problem of finding a minimal perturbation to the agents' outcomes that achieves both fairness and popularity, e.g. Program 5 for randomized classification. For a target population with feature vectors X, we shift individuals' outcomes f F (X) or expected outcomes E[f F (X)] by a perturbation vector p. For deterministic de-cisions, p 2 { 1, 0, 1} n , while for stochastic decisions p 2 [ 1, 1] n . The optimization goal in either case is to minimize kpk q for some `q-norm (q 2{ 1, 2, 1}) such that the final decisions, whether they involve predictions (f F (X)+p, or E[f F (X)]+p) or allocations (I(X,h,k)+p, or E ⇥ I(X,h,k) ⇤ + p) are both -fair and -popular. Since DOS does not use knowledge of true labels Y , it can be applied directly at prediction time to a population of individuals. However, this also means that it can only be applied when the measure of fairness is independent of the true labels Y (for example, ensuring equality of positive rates).
k-Quantile Lottery Shift (k-QLS) Another option for creating popular and fair classifiers is to directly minimize a loss function regularized by the distance of the fair-and-popular classifier from the fair classifier (distance is measured on predictions at training time), e.g. Program 12 for randomized classifiers. k-QLS-based postprocessing achieves this goal by partitioning scores h F (X) for a population X into k bins (based on quantiles). The goal is then to compute probabilities p (g) `for each bin `and group g, which minimize empirical risk and change to each agent's outcome, while achieving -popularity and -fairness. This is done at training time. Then at prediction time, we take all agents in group g with scores in bin `and run a lottery, where each agent is classified as 1 with probability p (g) `, and 0 otherwise. Since k-QLS is applied on the training dataset, it also allows us to use fairness metrics that depend on labels Y ; for this reason k-QLS is not used in allocation, where Y is typically unknown.
k-QLS is motivated by works such as (Hardt, Price, and Srebro 2016;Pleiss et al. 2017;Kamiran, Karim, and Zhang 2012;Canetti et al. 2019;Lohia et al. 2019) which aim to postprocess a conventional model to achieve -fairness by running an "inclusion/exclusion" lottery on groups of agents. However, k-QLS differs from these approaches: shifting all outcomes of a group, even in a randomized manner, is too granular to achieve -popularity, and thus we shift outcomes within k quantiles. In Section C.3 of the Supplement we demonstrate the poor performance of group level shifts compared the higher precision shifts of both the quantile shifts of k-QLS and the individual shifts of DOS.
Remark 3.1. Achieving -popularity and -fairness may be infeasible in general. However, for common efficacy metrics (e.g., PR, FPR, and TPR), doing so is always possible. Both DOS and k-QLS have a feasible solution for any level of -popularity and -fairness, for both randomized and deterministic models.
When the conventional model f C , and -fair model f F are deterministic, the optimization problems defined for both the DOS approach and the k-QLS approach can be efficiently solved for any U defined by an additive efficacy metric M. In both cases, since model decisions are binary, post processing amounts to finding some set of agents negatively classified by f C , which minimally impact loss while not violating fairness, when positively classified.
Theorem 3.2 (Informal). When classifiers produce deterministic outcomes and U is defined by an additive fairness metric, the optimization problems for both DOS and k-QLS can be solved in polynomial time.
We defer the formal statement of this claim, and a full discussion of deterministic postprocessing, to the Supplement.
Next we investigate popularity as it relates to randomized classifiers. Recall that in the case of randomized classifiers DOS aims to minimally shift the expected outcomes of f F on a population (X,G), with unknown true labels Y , to produce the -popular -fair model, which we denote by f P , where
Thus, DOS aims to solve the following optimization problem:
for q 2{ 1, 2, 1}. A key challenge is that the popularity constraint ( 7) is discrete and non-convex, amounting to a combinatorial problem of identifying a subset of |X| individuals who prefer the f P to its conventional counterpart f C . Nevertheless, this problem can be solved in polynomial time.
Algorithm 1: (Randomized DOS) Postprocessing technique for converting a -fair model f F into a -popular -fair model f P . Input: population: (X,Y,G), -fair model: f F , conventional model: f C , popularity: Result: weights p s.t. f P = f F + p is -popular andfair 1: G g := {i :
build Program 5 and replace Constraint 7 with S i 6:
p i = solution to the modified program 7: end for return p ⇤ = arg min i kp i k Theorem 3.3. Let f C and f F be respectively a conventional and -fair randomized classifier. Let U be derived from an additive efficacy metric M which is independent of Y (e.g., PR). Then for q 2{ 1, 2, 1} Program 5 can be solved in time Θ( nT) (where Θ(T ) is the time required to solve a linear program or semi-definite program, as appropriate) by Algorithm 1, which returns a -popular, -fair model
this holds for at least m = n agents then f P is -popular.
In the case of DOS postprocessing, if a specific set of m constraints is required to hold, rather than any m constraints, the problem is tractable as it is a linear program (q =1 , 1) or semi-definite program (q =2).
To order the set of possible constraints such that only a polynomial number must be examined, we make use of the following observations: for any two agents i, j 2 G g , 1.) since U is additive and independent of Y , unfairness is invariant under any change to p i ,p j which preserves p i + p j , and 2.) if
Thus, for any solution p where h C (x i ) h F (x i )+p i , but h C (x j ) > h F (x j )+p j , permuting p i and p j does not affect loss, fairness, or popularity, (when permutation is infeasible, shifting the maximum allowed weight from p i to p j is sufficient). Since the problem is invariant under such permutations, we need only consider imposing h
Thus, each G g can be ordered such that for i, j 2 G g , if
Since the intragroup decisions are made trivial via this ordering, only the intergroup decisions remain. Since at least m popularity constraints need to hold, and there are m ways to select exactly m total constraints between the two groups while preserving the intragroup ordering, there are only m sets of constraints that need investigation. Each set corresponds to solving either a LP or SDP which takes time Θ(T ) to solve. The specific running time of each program type is outlined in the Supplement. Thus the total running time of DOS is Θ( nT).
Next we turn our attention to resource allocation, in which k<nequally desirable resources are allocated to a population of size n. Recall that the randomized allocation scheme given by I(X,G) assigns resources to agents where E ⇥ I i (X,G) ⇤ 2 [0, 1] gives the probability that agent i receives a resource with allocation performed over population (X,G). For notational convenience, we use
to represent the probability that agent i receives the resources and suppress the expectation and implicit dependence on the population (X,G).
Scarce resource allocation is particularly well suited for DOS as true labels (with respect to the allocation decision) are typically unknown. In this case, DOS postprocessing is given by,
We now show that DOS in resource allocation settings remains tractable. Theorem 3.4. Let I C and I F be a conventional and -fair allocation scheme, respectively, and U be derived from an additive efficacy metric M which is independent of Y (e.g., PR). Then for q 2{ 1, 2, 1} Program 8 can be solved in time Θ( nT) by Algorithm 1 which returns a -popular,fair allocation if one exists.
Proof Sketch. In the case of scarce resources, agents can again be ordered in an identical fashion to the classification setting (Theorem 3.3). Note that for any solution p and any i, j 2 G g , the resource constraint P n i=1 I F (i)+p i k is invariant to any change in p i ,p j , which preserves p i +p j . Thus a similar argument to Theorem 3.3, with a few caveats relating to infeasible solutions, holds. Specifically, this yields n programs (either LPs or SDPs), each of which is solvable in time Θ(T ). Thus DOS post processing for resource allocation can be computed in time Θ( nT).
Finally, we explore k-QLS postprocessing for randomized classifiers. k-QLS creates k intervals by the quantiles of h F (X), where k is chosen by the model designer. Specifically, let ⇢ `be the maximum score associated with quantile `of h F (X). Each interval is given as
with the understanding that ⇢ 0 =0and ⇢ k =1 . On each interval I `, and for each group g, a parameter p
Finding the optimal lottery probabilities can formulated as the following optimization problem:
where L is expected training error. As was the case for DOS postprocessing with randomized classifiers, the constraint that fraction of the population prefers f P over f C is discrete and non-convex. Indeed, unlike DOS, the k-QLS problem becomes strongly NP-hard. Theorem 3.5. Postprocessing to achieve -popularity and -fairness with k-QLS (i.e., solving Program 12) is strongly NP-hard when models are randomized, and U is derived from an additive efficacy metric. We defer this proof to Section B.3 of the Supplement.
The intractability stems entirely from the model designer's ability to choose the number of quantiles k: if k is fixed, the problem can be solved in polynomial time as shown in the following theorem. In practice, we can fix k to be small, thus obtaining a tractable algorithm. Theorem 3.6. Let f C and f F be a conventional and a -fair randomized classifier respectively. Let U be derived from an additive efficacy metric M. Then for a fixed number of quantiles k, Program 12 for q = {1, 2, 1} can be solved in polynomial time, thus obtaining -popular -fair decisions.
Proof. As was the case for DOS applied to randomized classifiers, k-QLS applied to randomized classifiers is tractable if a specific set of m = n agents is required to prefer f P , rather than any m agents. When the number of intervals is constant it is straightforward to induce an ordering on agents which explores only a polynomial number of constraint sets. Specifically, let G (g,`) = {i 2 [n]:g i = g and h F (x i ) 2 I `}. Then agents in each G g can be ordered by the magnitude of p (g) `required such that they prefer f P to f C . Order G g such that for i, j 2 G g if i<jthen h C (x j ) h C (x i ), then if agent i 2 G g prefers f P to f C , so does every j i. There are 2k such sets, each containing at most n /k agents. Since the popularity over each G g can be parameterized by the identity of the agent with the largest value of h C (x) who prefers f P , there are no more than ( n) k unique values under this parameterization, and thus no more than ( n) k sets of constraints need be examined; each examination requires only polynomial time.
In this section we empirically investigate the relationship between popularity and fairness, and evaluate the efficacy of the proposed postprocessing algorithms. Each experiment is conducted on four data sets: 1) the Recidivism dataset, 2) the Income dataset, 3) the Community Crime dataset, and 4) the Law School dataset. In each dataset features can be continuous or categorical; each label is binary and defined such that 1 is always the more desirable outcome, e.g. in the Recidivism dataset y =1indicates not reoffending. A specific description of the label is given in the Supplement. Group membership is defined by race for Community Crime and Law School, and by gender for Recidivism and Income; either feature is assumed to be binary. All other sensitive features, such as age, are removed from the dataset. We consider three fair learning schemes: the Reductions algorithm (Agarwal et al. 2018), the CalEqOdds algorithm (Pleiss et al. 2017), the KDE algorithm (Cho, Hwang, and Suh 2020). Results for the latter two are provided in Section C of the supplement.
Popularity of Current Fair Learning Schemes: We begin by considering popularity of group-fair classifiers. The fractions of the overall population, and subgroup population, which prefer the fair classifier are shown in Figure 1, where fairness is achieved using the Reductions method.
Not surprisingly, we see that in all instances the disadvantaged group G 0 prefers f F at far higher rates than G 1 . With the exception of the CalEqOdds algorithm (which achieves fairness via group specific score shifts, resulting in far stronger group-level preference over classifiers), results for other methods are similar; these are provided in the supplement. Overall, randomized fair classifiers frequently have popularity of less than 50% . On the other hand, fair deterministic classifiers are relatively popular in most cases.
In either case, however, postprocessing can be used to further boost popularity of group-fair methods.