Equalizing Recourse across Groups

09/07/2019 ∙ by Vivek Gupta, et al. ∙ THE UNIVERSITY OF UTAH 28

The rise in machine learning-assisted decision-making has led to concerns about the fairness of the decisions and techniques to mitigate problems of discrimination. If a negative decision is made about an individual (denying a loan, rejecting an application for housing, and so on) justice dictates that we be able to ask how we might change circumstances to get a favorable decision the next time. Moreover, the ability to change circumstances (a better education, improved credentials) should not be limited to only those with access to expensive resources. In other words, recourse for negative decisions should be considered a desirable value that can be equalized across (demographically defined) groups. This paper describes how to build models that make accurate predictions while still ensuring that the penalties for a negative outcome do not disadvantage different groups disproportionately. We measure recourse as the distance of an individual from the decision boundary of a classifier. We then introduce a regularized objective to minimize the difference in recourse across groups. We explore linear settings and further extend recourse to non-linear settings as well as model-agnostic settings where the exact distance from boundary cannot be calculated. Our results show that we can successfully decrease the unfairness in recourse while maintaining classifier performance.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Algorithm-assisted decision making is now ubiquitous in domains as diverse as employment, education, credit, insurance, and criminal justice. How can we be sure that the decisions made through such systems will be fair and justifiable? One way in which we have traditionally placed checks on the power of decision-making systems, especially in settings where the decisions may have significant effects on our lives, is by providing the subject of the decision with a mechanism for recourse [USL19]: the ability to ask how the decision was made and what can be done to change it. This idea of recourse is a part of the Equal Credit Opportunity Act –an individual can demand that reasons for a loan denial be provided – and research suggests that it might also eventually be a component of the right to explanation implicated in the General Data Protection Regulation of the European Union. But the idea of recourse is more than a legal concept: it can be viewed as an ethical and just way to make decision-making more transparent. If a decision that materially affects us is made using factors that we cannot realistically change (for example a decision based on our gender or race), then this decision eliminates our agency and is therefore unjust [RCP]. Thus providing opportunities for recourse is a valuable goal independent of concerns about fairness and non-discrimination.

Recourse must be feasible to be useful. If an individual is rejected for a small business loan because of a low credit score and the recommended “recourse” is to attempt double their net worth, this is not a reasonable outcome. The situation is even worse if individuals from a different demographic are given a different recommendation – for example, to take an (inexpensive) course on money management for small businesses – that is easier to act on. In other words, providing an explanation is useful, but providing paths for recourse that are feasible and do not unduly burden disadvantaged groups is crucial to provide transparent opportunities for advancement.

Our Work.

In this paper, we build classifiers that maintain prediction accuracy while also ensuring opportunities for feasible recourse across groups. To do this, we define a general notion of recourse as the distance from the decision boundary (inspired by the work of [USL19]) and utilize it as a regularizer. In addition, we extend previous linear approaches for recourse [USL19] to non-linear settings, including both SVMs as well as models defined by black-box classifiers where we cannot explicitly compute the distance to the decision boundary. Our results show that we can successfully equalize recourse across groups in both SVM and model agnostic settings while maintaining accuracy.

On the ethical validity of equalized recourse. Recourse as a good is different from fairness/nondiscrimination. Regardless of the decision process, recourse captures the ethical notion of agency: that individuals should be able take actions to rectify their circumstances, and that decisions are not based on immutable factors. Recourse might not be cheap – but it should not be intrinsically infeasible. The idea of equalizing recourse across groups is then more than “if it’s good it should be good for all”. Rather, if we like recourse we don’t want a system to provide token recourse to one group whilst covertly discriminating. That is the justification for our work. We don’t claim that recourse must be universally applicable. Rather, it focuses on a specific kind of unfairness and seeks to address it when appropriate. We also strenuously push back on the idea that equalizing recourse (or even other fairness measures) “lowers the bar for the disadvantaged”. This perspective assumes that the bar was set correctly to begin with, rather than potentially being evidence of structural disadvantage.

2 Related work

The idea of recourse as a desirable property of a classifier was introduced and formalized by [USL19]

. In their work they analyze a linear classifier and show how to use integer linear programming to generate a set of actionable input variables (flipsets) to change the prediction for any individual. Their work can be thought of as part of a larger effort in fairness via

counterfactual analysis [WMR17] where the goal was to define fairness in terms of counterfactual properties of classifications. Strategic classification techniques [HMPW16, HIV19, MMDH19] which aim to modify the input in order to trick the classifier and achieve the desired decision are another concept related to recourse.

Our work fits within the larger setting of methods that try to find classifiers that satisfy some notion of fairness via constraints during the model learning process such as for demographic parity, equalized odds

[ZVRG17, ZVGRG17] and individual fairness [DHP12]. But recourse is orthogonal to those notions of fairness being studied. That is, it makes sense to talk about recourse even if the classification is unfair in other respects, because measures of fairness apply at the point of decision, whereas recourse is a post-decision analysis. Equalizing recourse affects distance to the boundary, but does not necessarily affect the decision itself. Having said that, studying the interaction between fair classification and recourse is a worthwhile topic for further exploration. Our approach in this paper involves shifting a baseline classifier through (iterated) processing which is similar to works such as [FKL16, GCGF16, DIKL18, PRW17, WGOS17]. Recourse can be viewed as a kind of explanation: an indicator of how to change one’s classification. We employ the well-known method LIME [RSG16, RSG18] to construct local explanations that guide our search for a classifier that equalizes recourse. [DCL18] recently proposed a method to provide minimal absent feature sets from examples which are sufficient for flipping their prediction.

3 Definitions

Recourse (as defined by [USL19]) is the cost of changing a decision from negative to positive. Formally, assume that we are given a classifier and a metric . The distance metric captures the “cost” of changing the features of a point , i.e., the effort involved in moving from to is .

Given a point such that , the recourse of is defined as . In other words, the recourse of a point is the (closest) distance from that point to the classifier.

The recourse of a set is the average recourse of all points in . Given a set of labeled pairs with a group membership function , let .

Problem 1 (Equalizing Recourse)

Given a set of labeled pairs and , find a classifier such that the loss is minimized under the constraint that where .

Note.

While we borrow the definition of recourse from [USL19], we do not also adopt their notion of a set of permissible actions to achieve recourse. This is motivated by the fact that we are trying to optimize recourse across groups rather than generate a flip set for a specific individual.

4 Optimizing for recourse

We now present our optimization frameworks for solving Problem 1. We start with the case where the classifier can be written explicitly.

4.1 Optimizing for recourse with an explicit classifier

For a classifier with a clear functional expression , we can write down expressions for the minimum distance of a point to the boundary of the classifier that is recourse . We then can simultaneously optimize the loss as well as the recourse.

Below, we show the optimization for equal recourse in kernelizable support vector machines; however, our method is quite general and can be transferred to other kernelizable classifiers such as Perceptron.

Support Vector Machines: Consider the case of support vector machines. For any linear classifier of the form , the distance to the decision boundary from a point is . In general, different features might be harder to modify, which can be modeled by scaling each dimension with a factor

. Effectively this amounts to applying a linear transformation

via a diagonal matrix where , and then computing the corresponding distance . If we now introduce a kernel , the above distance calculation must happen in the lifted feature space where is mapped to , and so the corresponding distance to the boundary is given by

The recourse of a set is merely the average recourse of the points in the set and the recourse difference .

SVM primal form. We formulate a data-driven primal-form constraint of kernel SVM inspired from [DC12] with an explicit recourse constraint.

such that
where

specifies a bound on recourse difference and controls the relative weight of in the optimization. are slack variables and is the slack for SVM classification.

SVM dual form. The dual form of the kernel SVM with implicit recourse constraints is the following:

(1)

where , , , and are dual variables.

We use the conditions with respect to and (equating gradient to 0) to find the optimal parameters.

(2)
(3)
(4)
(5)

Iterative Procedure: Since recourse is dependent on the prediction labels, that is which example is classified as negative, to calculate recourse for next iteration (), we can use the predictions of the last iteration() of SVM () when defining . Then, we can re-write as the following for the optimization,

(6)

where, denote the classification of the sample at iteration , and is the cardinality of set to which belongs.

Recourse Scaling: We scale the recourse by , similar to [DC12], because we care mostly about the relative difference between recourse. moving across the boundary requires traveling at least and so we can measure the difference in recourse in units of the margin by dividing by the margin which is equivalent to multiplying by .

Let’s denote = , C as the diagonal cost matrix, by substituting in equations (6), (2) and (3),

(7)
(8)
(9)

Equating equations 4 and 5 to zero cancels the terms involving and in the Lagrangian (equation 1).

(10)

We now simplify the terms in the Lagrangian to obtain 10 by equating equations 8, 9 to zero and substituting .

(11)

Substitute and , we get the following

(12)

Upon further simplification and cancelling the terms using equation 9, we obtain the following:

(13)

The final Lagrangian formulation after simplification would be:

(14)
s.t

We denote = , where is the classification of the sample at iteration , and is the cardinality of set to which belongs. Here, is the diagonal cost matrix.

To write the final Quadratic formulation, we need to introduce a pseudo recourse point (), where and , where, = , where is the cardinality of the set to which belongs. Let’s denote a new dual variable and new constraint constant .

Simplified Lagrangian dual formulation with the pseudo recourse point will be:

(15)

The final Quadratic form (QuadForm) for dual SVM is :

(16)

Here, has entries , , and , where , , and . Note that is a symmetric and positive semi definite matrix.

Here, we have variables , equality matrix , where inequality matrices and , where

is the identity matrix,

, and .

This form is compatible with any quadratic programming solver (CVXOPT) [ADV13] which results in the optimal values of , i.e., the optimal values of . Sometimes, solving a complete QP is intractable as it take an order . However, we can solve sub-problem to speed out the QP. We can use the standard SMO technique [ZYX08, KSBM01] which optimise the dual variable’s efficiently.

We can calculate by equating 8 to and substituting :

(17)

For a linear kernel will be a vector expressed in a closed form. However, for any other kernel, we cannot calculate explicitly because of the kernel; however, we can explicitly calculate as follows because of kernel property (obtained by applying and some simplification):

(18)

For calculating the bias , we need to use the support vectors, and those samples whose is significantly above a certain threshold should be on margins. To calculate the final value, we can take means of all values from support vectors:

(19)
(20)
(21)

The final classifier decision for a new prediction example can be obtained as follows:

(22)

Substitute from 17 into 22 yielded the final expression,

(23)

We can similarly calculate the final recourse ()) by substituting from 17 into 6 as follows,

(24)

where, from 18 , and the rest of terms are as defined before.

Here, in equations 23 and 24 we can use any suitable kernel, e.g., polynomial, linear, radial etc. to obtain the kernel matrix terms.

An iterative procedure

While the above formulation allows us to equalize recourse, it cannot be solved as written. This is because the specification of the dual requires the knowledge of which points are classified negatively due to the calculation of . However, for any fixed classifier (), we can identify the points that are classified negatively and use that to compute a new recourse-equalized classifier. This suggests an iterative strategy where the classifier from the previous round (t) i.e. is used to identify negatively-classified points which are then used to compute the classifier for this round (t+1) i.e. (). We summarize this in Algorithm 2.

An example.

We applied Algorithm 2 to two synthetic toy datasets to analyze our approach. In 1, we can see the motivation behind introducing recourse as a regularizer to be equalized across groups. We illustrate the positive class on the solid orange region with red datapoints. Conversely, the negative class is shown on the solid white background. The decision boundary is a solid black line that distinguishes the two areas. The dashed lines are the indicators of the margins. Our goal is to equalize recourse for the negatively classified groups in the brown transparent region that is closer to the boundary and the farther in the gray transparent region which has a larger recourse. As we increase and thus the importance of equalizing recourse, we observe that in the linear case, the decision boundary tilts such that the under-represented negative group will get closer to the decision boundary and the recourse is equalized. In the non-linear case, the boundary margins (dashed lines) are uni-modal around the positive class (brown), thus making it farther from the outer negative class (gray). As we increase , the boundary margins (dashed lines) become bi-modal with two dashed margins around both groups which means recourse is equalized.

Figure 1: Equalizing recourse across groups, from left (vanilla SVM ) to right ( increased).

Flipset generation. Similar to [USL19], we can generate flipsets by finding the closest points to the classifier boundary according to our notion of recourse. This is similar to the linear programming optimization approach proposed by [USL19]. One difference is since our attributes are continuous, we have more freedom in finding the flipsets since we know the exact closest boundary points according to our recourse definition. Besides, if we have constraints on attributes (e.g., some features are immutable and cannot be changed such as gender) we can use the ILP/LP idea used in [USL19] and utilize it in our balanced recourse classifier to obtain the required flipsets. Thus, our approach is orthogonal to [USL19]’s ILP/LP and can generate flipsets.

One might think of extending the ILP-based formulation [USL19] for obtaining fairness in recourse on top of generating flipsets without changing the boundary. For example, to achieve equal recourse one could optimize using linear programming for finding flip-sets which ensure equal recourse. Here, is the cost of changing features of to achieve recourse using the action set as defined in [USL19]. However, this idea contradicts the original notion of minimum actionable flipsets, which is achieved by minimizing . The flipset solution for equal recourse might not necessarily be of the minimum cost. So there might be other low-cost flipsets for actionable recourse. Thus, the individuals always have the right to choose the other flipsets that have a lower cost for them compared to the ones that remove the unfairness. Hence, changing the boundary is essential for yielding equal recourse across groups. In our case, we are changing the boundary to ensure that the minimum flipsets yields equal recourse.

4.2 Model agnostic settings

The previous approach for equalizing recourse is only applicable where we have an explicit mathematical formulation of the distance of a point from a classifier boundary and also we know the full objective of the underlying model. This is not the case with any arbitrary non-linear classifier. Also this scheme is not feasible in the model agnostic setting with black-box access. Moreover, finding a closed-form formula of the minimum distance of a point from a non-linear boundary (hence recourse) is either hard or infeasible in general.

To remedy this situation, we propose an approach that is based on LIME, a technique introduced by [RSG16]

to provide a local explanation for black-box classifiers. LIME works by sampling points from a normal distribution centered at a particular point or the mean of the dataset. LIME then fits a linear model using weighted regression for the sampled points against the decision produced by a black-box classifier. Our method hinges on the idea that LIME lets us approximate a complex decision boundary by training a linear classifier for every individual datapoint, and therefore, allowing us to compute the approximate distance of a point from the actual non-linear decision boundary. Once we have distances (and therefore the estimates for recourse

for each group), we can build a new classifier. However, since the classifier is a black-box we cannot explicitly introduce a regularization term as before.

Therefore, to change the black box non-linear boundary to equalize recourse, we propose an example-dependent weighted learning approach. The idea is that we can retrain the black-box classifier by passing the weight parameters for each sample point. Intuitively, we will re-weight points so that groups with large recourse (i.e., a larger distance to the decision boundary) are given a lower weight than groups with small recourse. The effect of this is to move the decision boundary away from the highly weighted points and towards the lower-weighted points, which is the desired effect.

Algorithm.

At first, we generate a set of samples (neighborhood) using multivariate normal centered at the mean of the dataset and scaled accordingly. We then train the black-box classifier. For every datapoint in the training set 10 most important features are chosen using a weighted regression using an exponential kernel with Euclidean distance metric for the sample weights. Now finally a ridge linear classifier is trained on the samples using only the selected features. This classifier is used to calculate the approximate distance of the datapoint from the original black-box classifiers boundary. We do this for all the training points from the trained classifier’s boundary. And we average it over several sets of samples normalized by the difference between maximum and minimum distances for the negative class. We keep the weights of points in the positive class (), unchanged (i.e. 1). To calculate the weight of point () in the negative class (), we estimate the normalized minimum average distance (namd()) of from the decision boundary

Here is set of negatively classified points, We can set the weight of any point by using the following formula,

(25)

Now, we retrain the classifier with these new sample weights. The classifier tries to be more certain in predicting the class of a point which is weighed more than the ones that are weighed less. Therefore, the above weighing scheme shifts the classifier’s boundary closer to the negative points that are farther (in other words have smaller weights) while maintaining the distances of the closer ones and the positive class mostly unchanged. Refer to Algorithm 2 for a pseudo-code description.

Data: , , , g,
Result: , and
/* Initialization by Vanilla SVM */
1 ; = ;
2 QuadForm(, , , g, , );
3 Update = h();
/* SVM with a Recourse Constraint */
4 for  do
5       QuadForm(, );
6       Update = h();
7      
8 end for
return
Data: , ,
Result:
/* Choose best LIME fit */
1 Choose best LIME samples in terms of prediction accuracy.;
2 ;
3 Train classifier () with samples weights ();
4 Find average recourse against using ;
5 Update sample weights () using Eq.25;
6 Train classifier () with sample weights ();
7 Find average recourse against using ;
return
(a) Recourse regularized SVM classification
(b) Recourse equalization in black-box classification
Figure 2: Algorithms for Recourse Equalization a) Regularised SVM and b) Black-Box Classification
(a) Recourse regularized SVM classification

5 Experiments

In this section, we will provide the details of our datasets and the experimental settings.

Dataset description. We evaluate both of our approaches on 4 datasets, credit [YL09], german [BL13], givemecredit and propublica.222german: https://github.com/algofairness/fairness-comparison/tree/master/fairness/data/preprocessed, givemecredit: https://www.kaggle.com/c/GiveMeSomeCredit, propublica: https://github.com/propublica/compas-analysis. We set the target variable as binary with labels +1 and -1 for all datasets. Positive sensitive attribute for credit was chosen to be Married and everything else as negative, the target value is -1 if the person defaults on a future credit card payment (we use a preprocessed version from [USL19] that uses the financial features). In the case of givemecredit, after removing rows with missing information, the negative sensitive class was individuals with age under 35 and target is if the individual will be in financial distress in the next two years. We used gender as sensitive attribute and set female as the negative class for both german and propublica. The target attribute for german is credit risk of a person and for propublica it is whether an individual would recommit a crime within 2 years from release.

Experimental settings.

For every run, we choose a random sample from a dataset. We choose a random sample out of all 4 datasets which is datapoints for credit (13 features), and for givemecredit (10 features), and propublica (402 features) and german (59 features). We then use a split for train and test respectively. We perform 10-fold cross-validation in terms of minimizing recourse difference to choose the best parameter set for both linear and non-linear kernels. We search over and polynomial kernel with . We then perform 10 iterations of our SVM-based algorithm using the optimum parameter set to obtain the final classifier. The costs for all the features are set to 1, and (cost parameter) is set to 10. We use the CVXOPT package [ADV13] for solving the quadratic formulation, shown in 16.

In the agnostic setting, we run our experiments with datapoints for all datasets using the following parameters. We execute 5 runs to choose 2 best set of samples for approximating black box classifier by LIME in terms of accuracy. We execute 10 independent runs of algorithm 2

and aggregate the results to generate the plots for all 4 datasets. We present our results on three black-box classifiers, namely Random Forest, Logistic Regression, and AdaBoost from the scikit-learn (v0.21) package

[PVG11].333Specifically, sklearn.ensemble.RandomForestClassifier, sklearn.linear_model. LogisticRegression, sklearn.ensemble.AdaBoostClassifier We choose both Adaboost and Random forest primarily because they are non-linear in nature and lack the closed-form mathematical objective function that our previous method was based on. Our reason for choosing Logistic classifier is mainly to check how the black-box strategy works for simple linear classifiers. We use the Logistic and Adaboost classifier with default parameters and RandomeForest is trained with max-depth (tree depth) set to 4 to avoid over-fitting and everything else is set as default.

In the SVM and model agnostic settings, we evaluate our proposed approaches for equalizing recourse before and after

regularization and sample re-weighting, respectively. All results are reported with box and whisker plots for complete details on the mean, median, and the percentiles of the accuracy and the recourse difference distributions. The outliers are not shown in the plot; however, they are considered in the computations.

Figure 3: Linear (upper) and polynomial (lower) kernels’ results on recourse and accuracy. Yellow circle is the mean, and the orange line is the median.

5.1 SVM results

We start by discussing the results of applying Algorithm 2, listed in Table 1. We can summarize these results by showing the percentage reduction in recourse difference () for different data sets (mean of the distributions). For this summary, we focus on the polynomial and linear kernels.

Linear Polynomial
Dataset Train Test Train Test
german 80% 0% 91% 23%
credit 52% 29% 51% 50%
propublica 84% 16% 99% 23%
givemecredit 90% 9% 100% 55%
Table 1: Summary of percentage reduction in recourse difference for polynomial and linear kernels

In general, we observe that we are yielding an improvement in the unfairness of recourse across groups in terms of the means of the difference. We also observe that the medians are decreasing (and we have smaller ranges of percentile distributions after applying our algorithm), which means at least around half of the runs are yielding a smaller recourse difference compared to before even though the mean might remain the same due to outliers, e.g., in german the test recourse difference is not decreasing significantly ‘after’ applying our algorithm compared to other datasets. However, more than (e.g., lower th percentile for both kernels) of the runs have a smaller recourse difference after according to the median, or up to th percentile of givemecredit

data have a smaller recourse. In addition, we observe that the accuracy roughly stays the same (it decreases no more than 1% and sometimes even slightly increases). Furthermore, the improvement we acquired using the linear kernel is less compared to polynomial, we assume this is due to fewer degrees of freedom (fewer model parameters) in linear kernel compared to polynomial to change the decision boundary.

Figure 4: RandomForest (left), LogisticRegression (middle), AdaBoost (right) classifier’s results on recourse and accuracy. Yellow circle is the mean and the orange line is the median.

5.2 Model agnostic results

We have listed the results of applying Algorithm 2 in Table 2. Similar to SVM, both train and test sets show improvements in terms of the mean, median and percentile distributions of the recourse difference before and after equalization. In addition, the mean accuracies for credit, german, and propublica are not reduced by more than 2%, except for the training accuracy during logistic regression for german (% drop), test/train accuracies for propublica (around 7% and 5% drops, respectively), and test/train accuracies for givemecredit (around 7% for both).

General Analysis. Overall, these results show that our proposition to define recourse for a group as the mean of the distances from the decision boundary and its incorporation in the SVM formulation as a regularizer can reduce the unfairness in the recourse across groups. Additionally, the simple re-weighting strategy works well in terms of recourse difference reduction for classifiers that support sample weights and could be useful for black-box scenarios where more fine-grained control cannot be obtained for explicit distances from the boundary in terms of model parameters. Both methods yield these results without affecting the accuracy significantly. In addition, the median and percentage-based distributions of the recourse difference across various runs decrease after equalization.

Random Forest Logistic Regression AdaBoost
Dataset Train Test Train Test Train Test
credit 41% 99% 90% 100% 83% 95%
german 64% 42% 8% 53% 37% 60%
propublica 59% 43% 29% 52% 49% 56%
givemecredit 83% 70% 20% 100% 66% 14%
Table 2: Summary of percentage reduction in recourse difference for Model Agnostic

6 Conclusion and future work

In summary, this work introduces a new notion of fairness in recourse (Equalizing Recourse), i.e., classifiers that maintain a good performance while providing opportunities for feasible recourse across groups. We proposed various approaches based on a general definition of recourse as the absolute value of the difference between the average of the distance of all points from the boundary in a group. We then utilized this definition to make classification models fair in terms of recourse. Our approach is generalizable to linear and non-linear settings, including kernelizable (support vector machines) and non-kernelizable (decision trees) models. We empirically validated these methods through intensive experiments. This is the first work on outcome-independent fairness in terms of recourse (equal costs for changing the outcome) as far as our knowledge. Incorporating multiple groups/sensitive attributes is a possible future direction. In addition, instead of only equalizing the means of the distributions, we can equalize the distributions themselves using correlation or KL divergence.

References

  • [ADV13] M Andersen, Joachim Dahl, and Lieven Vandenberghe. Cvxopt: A python package for convex optimization. abel. ee. ucla. edu/cvxopt, 2013.
  • [BL13] Kevin Bache and Moshe Lichman. Uci machine learning repository, 2013.
  • [DC12] Sune Darkner and Line H Clemmensen. Data driven constraints for the svm. In International Workshop on Machine Learning in Medical Imaging, pages 70–77. Springer, 2012.
  • [DCL18] Amit Dhurandhar, Pin-Yu Chen, Ronny Luss, Chun-Chen Tu, Paishun Ting, Karthikeyan Shanmugam, and Payel Das. Explanations based on the missing: Towards contrastive explanations with pertinent negatives. In Advances in Neural Information Processing Systems, pages 592–603, 2018.
  • [DHP12] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214–226. ACM, 2012.
  • [DIKL18] Cynthia Dwork, Nicole Immorlica, Adam Tauman Kalai, and Max Leiserson. Decoupled classifiers for group-fair and efficient machine learning. In Conference on Fairness, Accountability and Transparency, pages 119–133, 2018.
  • [FKL16] Benjamin Fish, Jeremy Kun, and Ádám D Lelkes. A confidence-based approach for balancing fairness and accuracy. In Proceedings of the 2016 SIAM International Conference on Data Mining, pages 144–152. SIAM, 2016.
  • [GCGF16] Gabriel Goh, Andrew Cotter, Maya Gupta, and Michael P Friedlander. Satisfying real-world goals with dataset constraints. In Advances in Neural Information Processing Systems, pages 2415–2423, 2016.
  • [HIV19] Lily Hu, Nicole Immorlica, and Jennifer Wortman Vaughan. The disparate effects of strategic manipulation. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 259–268. ACM, 2019.
  • [HMPW16] Moritz Hardt, Nimrod Megiddo, Christos Papadimitriou, and Mary Wootters. Strategic classification. In Proceedings of the 2016 ACM conference on innovations in theoretical computer science, pages 111–122. ACM, 2016.
  • [KSBM01] S. Sathiya Keerthi, Shirish Krishnaj Shevade, Chiranjib Bhattacharyya, and Karuturi Radha Krishna Murthy. Improvements to platt’s smo algorithm for svm classifier design. Neural computation, 13(3):637–649, 2001.
  • [MMDH19] Smitha Milli, John Miller, Anca Dragan, and Moritz Hardt. The social cost of strategic classification. In Proceedings of ACM FAT*, 2019.
  • [PRW17] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger. On fairness and calibration. In Advances in Neural Information Processing Systems, pages 5680–5689, 2017.
  • [PVG11] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
  • [RCP] Alan Rubel, Clinton Castro, and Adam Pham. Algorithms, bias, and the importance of agency. Jo Bates Paul D. Clough Robert Jäschke, page 9.
  • [RSG16] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144. ACM, 2016.
  • [RSG18] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Anchors: High-precision model-agnostic explanations. In

    Thirty-Second AAAI Conference on Artificial Intelligence

    , 2018.
  • [USL19] Berk Ustun, Alexander Spangher, and Yang Liu. Actionable recourse in linear classification. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 10–19. ACM, 2019.
  • [WGOS17] Blake Woodworth, Suriya Gunasekar, Mesrob I Ohannessian, and Nathan Srebro. Learning non-discriminatory predictors. arXiv preprint arXiv:1702.06081, 2017.
  • [WMR17] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 31(2):2018, 2017.
  • [YL09] I-Cheng Yeh and Che-hui Lien.

    The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients.

    Expert Systems with Applications, 36(2):2473–2480, 2009.
  • [ZVGRG17] Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P Gummadi. Fairness beyond disparate treatment & disparate impact: Learning classification without disparate mistreatment. In Proceedings of the 26th International Conference on World Wide Web, pages 1171–1180. International World Wide Web Conferences Steering Committee, 2017.
  • [ZVRG17] Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rogriguez, and Krishna P Gummadi. Fairness constraints: Mechanisms for fair classification. In Artificial Intelligence and Statistics, pages 962–970, 2017.
  • [ZYX08] Zhi-Qiang Zeng, Hong-Bin Yu, Hua-Rong Xu, Yan-Qi Xie, and Ji Gao. Fast training support vector machines using parallel sequential minimal optimization. In

    2008 3rd international conference on intelligent system and knowledge engineering

    , volume 1, pages 997–1001. IEEE, 2008.