Back

Fairness Constraints: Mechanisms for Fair Classification

Muhammad Bilal ZafarIsabel ValeraManuel Gomez RodriguezKrishna P. Gummadi

article2015AISTATS1,370 citations

Presents a flexible framework that integrates intuitive fairness constraints directly into the training of logistic regression and support vector machines, allowing practitioners to precisely balance accuracy and demographic equity without severe performance losses.

Listen

Automated decision-making systems increasingly govern critical life outcomes, including loan underwriting, criminal justice risk assessments, and hiring. However, models trained on historical data often perpetuate or amplify discrimination against protected demographic groups. Legal standards evaluate discrimination using two concepts: disparate treatment (making decisions explicitly based on sensitive attributes such as race or gender) and disparate impact (producing disproportionately negative outcomes for protected groups). Preventing both simultaneously is challenging because simply omitting sensitive attributes during training fails to prevent indirect bias, while using sensitive attributes during deployment can trigger unlawful disparate treatment.

The article introduces and evaluates a flexible mathematical framework designed to train fair classification models—specifically convex margin-based classifiers such as logistic regression and support vector machines. The objective is to simultaneously eliminate disparate treatment and control disparate impact, while also accommodating the legal business necessity clause, which permits a controlled degree of disparate impact to satisfy core performance requirements.

Directly enforcing legal fairness metrics—such as the standard 80%-rule or p%-rule—in machine learning models creates intractable, non-convex optimization problems. To overcome this, the authors develop an efficient convex proxy: the covariance between sensitive attributes and the signed distance of data points from the model's decision boundary. Using this proxy, the authors design two complementary model training formulations: one that maximizes accuracy under specific fairness constraints, and another that maximizes fairness under specified accuracy or loss thresholds. The approach was validated through synthetic experiments and empirical evaluations on two benchmark datasets: the Adult income dataset (45,222 records) and the Bank marketing dataset (41,188 records), evaluating binary, non-binary, and multiple simultaneous sensitive attributes.

The evaluation produced several key findings. First, the proposed fairness constraints enable precise, fine-grained control over the fairness-accuracy trade-off, achieving high levels of fairness (approaching a 100%-rule) with only minor reductions in overall accuracy. Second, unlike existing fair-learning benchmarks that require sensitive attributes at the decision stage, the proposed method removes sensitive attributes entirely during deployment, avoiding disparate treatment while matching or exceeding the accuracy of competing techniques. Third, the framework successfully scales to complex settings, including multi-category sensitive attributes (such as race) and multiple sensitive attributes simultaneously (such as gender and race combined), where alternative methods fail. Fourth, under the business necessity formulation, incorporating fine-grained individual loss constraints prevents previously approved candidates from being erroneously reclassified into negative outcomes when adjusting for fairness.

These findings demonstrate that organizations can deploy legally defensible automated classifiers that mitigate systemic historical biases without suffering significant operational or financial performance penalties. The dual formulations provide decision-makers with the flexibility either to meet strict regulatory non-discrimination thresholds or to fulfill performance-critical business objectives while minimizing residual bias.

Organizations developing automated screening or scoring pipelines should adopt boundary-covariance constraints within their classification workflows to achieve proactive compliance. Leaders should evaluate whether their operating environment prioritizes strict regulatory compliance (maximizing accuracy under fairness constraints) or operational performance (maximizing fairness under accuracy constraints). Further technical development is recommended to extend these constraints to broader problem domains, including regression, ranking, and recommender systems, and to analytically map covariance thresholds directly to exact p%-rule targets.

Confidence in these findings is high for standard convex classification models operating on historical tabular data. However, decision-makers should note that disparate impact mitigation assumes the ground-truth historical labels are inherently biased; in operational domains where unbiased ground truth is verifiable, alternative criteria such as disparate mistreatment may be more appropriate.

Max Planck Institute for Software Systemsasian continental ancestry grouplogistic regressionlogisticsfairness constraintsfair classificationdisparate impact
arXiv: 1507.05259
  • Paper: Fairness through awareness, Cynthia Dwork et al. (2012). This foundational paper formalizes algorithmic fairness constraints and statistical parity definitions that the source directly builds upon and incorporates into decision boundary optimization.
  • Paper: Learning Fair Representations, Richard Zemel et al. (2013). It establishes key representation learning techniques to eliminate protected attribute discrimination, providing essential context for in-processing fair classification mechanisms.
  • Paper: Certifying and Removing Disparate Impact, Michael Feldman et al. (2014). It defines formal mathematical criteria and data repair methods for disparate impact, serving as a direct prerequisite for formulating fair classification constraints.
  • Paper: A training algorithm for optimal margin classifiers, Bernhard E. Boser et al. (1992). It provides the foundational optimization formulation for support vector machine decision boundaries that the source extends with explicit fairness constraints.
  • Paper: The foundations of cost-sensitive learning, Charles Elkan (2001). It introduces the theoretical foundations of constrained and cost-sensitive classification that underpin trade-offs between decision accuracy and auxiliary objectives.
Cover for Fairness Constraints: Mechanisms for Fair Classification

Abstract

Algorithmic decision making systems are ubiquitous across a wide variety of online as well as offline services. These systems rely on complex learning methods and vast amounts of data to optimize the service functionality, satisfaction of the end user and profitability. However, there is a growing concern that these automated decisions can lead, even in the absence of intent, to a lack of fairness, i.e., their outcomes can disproportionately hurt (or, benefit) particular groups of people sharing one or more sensitive attributes (e.g., race, sex). In this paper, we introduce a flexible mechanism to design fair classifiers by leveraging a novel intuitive measure of decision boundary (un)fairness. We instantiate this mechanism with two well-known classifiers, logistic regression and support vector machines, and show on real-world data that our mechanism allows for a fine-grained control on the degree of fairness, often at a small cost in terms of accuracy.

Topics

asian continental ancestry grouplogistic regressionlogisticssensitive attributesupport vectorclassification algorithmclassification modelssvmadaptive boostingrandom forests

Keywords

fairness constraintsfair classificationdisparate impactdisparate mistreatmentfair decision boundarysensitive attributesconvex margin-based classifiersindirect discrimination

Table of Contents

  • 1 INTRODUCTION
  • 2 FAIRNESS IN CLASSIFICATION
  • 2.1 Fairness Definition
  • 3 OUR APPROACH
  • 3.1 Decision Boundary Covariance
  • 3.2 Maximizing Accuracy Under Fairness Constraints
  • 3.3 Maximizing Fairness Under Accuracy Constraints
  • 4 EVALUATION
  • 4.1 Experiments on Synthetic Data
  • 4.2 Experiments on Real Data
  • 5 DISCUSSION & FUTURE WORK
  • References
  • A Particularizing the Fairness Constraints for SVM
  • B Additional Experiments
  • B.1 Experiments on Non-linear Synthetic Data
  • B.2 Experiments on Real Data

Knowls

  1. Knowl 1 — Decision Boundary Covariance as a Fairness Metric

    definition

    In binary classification tasks with feature vectors x∈Rdx \in \mathbb{R}^d and sensitive attributes zz, the decision boundary covariance is defined as the covariance between the sensitive attribute values and the signed distance from the feature vectors to the decision boundary dθ(x)d_\theta(x):

    Cov(z,dθ(x))=E[(z−zˉ)dθ(x)]−E[(z−zˉ)]dˉθ(x)≈1N∑i=1N(zi−zˉ)dθ(xi)\text{Cov}(z, d_\theta(x)) = \mathbb{E}[(z - \bar{z}) d_\theta(x)] - \mathbb{E}[(z - \bar{z})] \bar{d}_\theta(x) \approx \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) d_\theta(x_i)

    where zˉ=1N∑i=1Nzi\bar{z} = \frac{1}{N}\sum_{i=1}^N z_i is the empirical mean of the sensitive attribute over NN training examples, and E[(z−zˉ)]dˉθ(x)=0\mathbb{E}[(z - \bar{z})] \bar{d}_\theta(x) = 0.

    For linear classifiers with parameter vector θ\theta (where the bias is absorbed into θ\theta by appending 11 to xx), the signed distance is dθ(xi)=θTxid_\theta(x_i) = \theta^T x_i, yielding the empirical covariance:

    Cov(z,dθ(x))≈1N∑i=1N(zi−zˉ)θTxi\text{Cov}(z, d_\theta(x)) \approx \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) \theta^T x_i

    Because dθ(xi)d_\theta(x_i) is convex with respect to θ\theta for convex margin-based classifiers, this covariance measure is convex in θ\theta. If a classifier satisfies demographic parity (the 100%100\%-rule where P(dθ(x)≥0∣z=0)=P(dθ(x)≥0∣z=1)P(d_\theta(x) \ge 0 \mid z=0) = P(d_\theta(x) \ge 0 \mid z=1)), the empirical covariance is approximately zero for large NN.

  2. Knowl 2 — Maximizing Accuracy Subject to Decision Boundary Fairness Constraints

    model/method

    To train a margin-based classifier that avoids disparate impact while maximizing accuracy, the classifier parameters θ\theta are obtained by minimizing a convex empirical loss L(θ)L(\theta) subject to two-sided constraints on the decision boundary covariance:

    \min_{\theta} \quad & L(\theta) \\ \text{subject to} \quad & \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) d_\theta(x_i) \le c \\ & \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) d_\theta(x_i) \ge -c \end{aligned}$$ where $c \ge 0$ is a covariance threshold chosen by the practitioner. As $c \to 0$, the decision boundary enforces greater demographic parity ($p\%$-rule compliance). Because both the objective and the fairness constraints are convex, every pair of covariance and empirical loss values on the optimization frontier is Pareto optimal. Disparate treatment is avoided at inference time because the classification decision $f_\theta(x) = \text{sign}(d_\theta(x))$ depends only on non-sensitive features $x$, whereas the sensitive attribute $z$ is used strictly during training to constrain the parameter search.

  3. Knowl 3 — Fair Logistic Regression with Margin Covariance Constraints

    equation

    For logistic regression, where class conditional probabilities are parameterized by p(yi=1∣xi,θ)=11+e−θTxip(y_i = 1 \mid x_i, \theta) = \frac{1}{1 + e^{-\theta^T x_i}} for labels yi∈{−1,1}y_i \in \{-1, 1\}, the accuracy-maximizing fair formulation corresponds to the constrained negative log-likelihood minimization:

    \min_{\theta} \quad & -\sum_{i=1}^N \log p(y_i \mid x_i, \theta) \\ \text{subject to} \quad & -c \le \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) \theta^T x_i \le c \end{aligned}$$ where $\theta \in \mathbb{R}^d$ includes the bias term (via an augmented feature $x_i$ with a constant $1$), $z_i$ is the sensitive attribute for the $i$-th instance, $\bar{z}$ is its sample mean, and $c \ge 0$ is the fairness covariance upper bound.

  4. Knowl 4 — Fair Linear and Kernel Support Vector Machines

    model/method

    The fairness covariance constraints can be embedded directly into linear and non-linear Support Vector Machines (SVMs).

    Linear SVM (Primal):

    \min_{\theta, \xi} \quad & \|\theta\|^2 + C \sum_{i=1}^N \xi_i \\ \text{subject to} \quad & y_i \theta^T x_i \ge 1 - \xi_i, \quad \forall i \in \{1, \dots, N\} \\ & \xi_i \ge 0, \quad \forall i \in \{1, \dots, N\} \\ & -c \le \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) \theta^T x_i \le c \end{aligned}$$ where $\|\theta\|^2$ maximizes the margin, $C$ penalizes margin violations $\xi_i$, and $c$ bounds the decision boundary covariance. **Nonlinear SVM (Dual with Kernel Trick):** For a non-linear mapping $\Phi(x)$ with kernel function $k(x_i, x_j) = \langle \Phi(x_i), \Phi(x_j) \rangle$, the signed distance is represented in the dual as $g_\alpha(x_i) = \sum_{j=1}^N \alpha_j y_j k(x_i, x_j)$. The optimization problem in dual variables $\alpha \in \mathbb{R}^N$ is: $$\begin{aligned} \min_{\alpha} \quad & \sum_{i=1}^N \alpha_i + \sum_{i=1}^N \alpha_i y_i \left( g_\alpha(x_i) + h_\alpha(x_i) \right) \\ \text{subject to} \quad & \alpha_i \ge 0, \quad \forall i \in \{1, \dots, N\} \\ & \sum_{i=1}^N \alpha_i y_i = 0 \\ & -c \le \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) g_\alpha(x_i) \le c \end{aligned}$$ where $h_\alpha(x_i) = \frac{\alpha_i y_i}{C}$.

  5. Knowl 5 — Maximizing Fairness under Global and Fine-Grained Accuracy Constraints

    model/method

    To implement the business necessity clause of disparate impact—where a classifier must achieve the maximum possible fairness subject to explicit performance thresholds—the training problem is inverted to minimize decision boundary covariance subject to loss bounds.

    Global Loss Constraint:

    \min_{\theta} \quad & \left| \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) d_\theta(x_i) \right| \\ \text{subject to} \quad & L(\theta) \le (1 + \gamma) L(\theta^*) \end{aligned}$$ where $\theta^*$ is the optimal unconstrained parameter vector, $L(\theta^*)$ is its training loss, and $\gamma \ge 0$ represents the maximum allowable proportional increase in overall training loss. **Fine-Grained Individual Loss Constraints:** When individual loss is additive across samples ($L(\theta) = \sum_{i=1}^N L_i(\theta)$), individual bounds can prevent desirable candidates from being misclassified to satisfy demographic quotas: $$\begin{aligned} \min_{\theta} \quad & \left| \frac{1}{N} \sum_{i=1}^N (z_i - \bar{z}) \theta^T x_i \right| \\ \text{subject to} \quad & L_i(\theta) \le (1 + \gamma_i) L_i(\theta^*), \quad \forall i \in \{1, \dots, N\} \end{aligned}$$ Setting $\gamma_i = 0$ for individuals classified as positive under $\theta^*$ guarantees that the fair boundary will not flip them to negative, causing the decision boundary to translate and rotate rather than purely rotate.

  6. Knowl 6 — Fairness Constraints for Multiple and Polyvalent Sensitive Attributes

    model/method

    The decision boundary covariance mechanism extends to settings with non-binary (polyvalent) attributes (such as race with K>2K > 2 categories) and settings with multiple simultaneous sensitive attributes (such as gender and race).

    For each sensitive feature dimension or one-hot encoded category k∈{1,…,K}k \in \{1, \dots, K\}, an unconstrained classifier parameter θ∗\theta^* is first trained to compute its baseline covariance:

    ck∗=∣1N∑i=1N(zi,k−zˉk)dθ∗(xi)∣c_k^* = \left| \frac{1}{N} \sum_{i=1}^N (z_{i,k} - \bar{z}_k) d_{\theta^*}(x_i) \right|

    A single multiplicative covariance relaxation factor a∈[0,1]a \in [0, 1] is selected, and simultaneous constraints are enforced for all kk:

    −ack∗≤1N∑i=1N(zi,k−zˉk)dθ(xi)≤ack∗,∀k∈{1,…,K}-a c_k^* \le \frac{1}{N} \sum_{i=1}^N (z_{i,k} - \bar{z}_k) d_\theta(x_i) \le a c_k^*, \quad \forall k \in \{1, \dots, K\}

    Decreasing aa toward 0 progressively eliminates disparate impact across all protected groups and subgroups simultaneously.

  7. Knowl 7 — Accuracy vs. Fairness Trade-Offs on Adult and Bank Marketing Benchmarks

    empirical result

    The covariance-constrained logistic regression (C-LR) and SVM (C-SVM) frameworks were evaluated on the UCI Adult Income dataset (45,222 instances, predicting income >50K>50\text{K} USD with gender and race as sensitive attributes) and the UCI Bank Marketing dataset (41,188 instances, predicting term deposit subscriptions with age as sensitive attribute), split 70% train and 30% test across 5 runs.

    Key empirical findings:

    1. On the Adult dataset (gender fairness), C-LR and C-SVM reduced empirical covariance to 0 and increased the p%p\%-rule from 24%24\% up to ≈100%\approx 100\% while accuracy dropped by only ≈4%\approx 4\% (from ≈84%\approx 84\% down to ≈80%\approx 80\%).
    2. Compared against Preferential Sampling (PS-LR, PS-SVM), which pre-processes training data, C-LR and C-SVM achieved strictly superior accuracy across all fairness levels; preferential sampling failed to reach an 80%80\%-rule on both datasets.
    3. Compared against Prejudice Remover Regularized Logistic Regression (R-LR), C-LR achieved comparable or slightly better accuracy-fairness trade-offs on Adult, while crucially maintaining compliance with disparate treatment laws because R-LR requires access to sensitive attributes during test-time inference.

  8. Knowl 8 — Global vs. Fine-Grained Loss Constraints Behavior

    empirical result

    When maximizing fairness under accuracy constraints on synthetic and real datasets:

    1. Global loss constraints (γ\gamma-LR) satisfy higher p%p\%-rules by rotating the decision boundary, which simultaneously demotes non-protected individuals from the positive to the negative class and promotes protected individuals from the negative to the positive class.
    2. Fine-grained loss constraints (Fine-γ\gamma-LR) with γi=0\gamma_i = 0 for non-protected individuals previously classified as positive prevent any positive non-protected instance from being demoted. The optimization satisfies fairness by shifting and translating the boundary to classify additional protected individuals as positive.
    3. As a consequence of preserving positive predictions for existing qualified subjects, Fine-γ\gamma-LR incurs a steeper reduction in test accuracy for the same target p%p\%-rule than γ\gamma-LR.

  9. Knowl 9 — Demographic and Outcome Distributions of the Adult and Bank Datasets

    data/table

    The evaluation uses two standard discrimination benchmarks showing strong historical label disparities across demographic groups:

    Adult Dataset (Gender) y≤50Ky \le 50\text{K} y>50Ky > 50\text{K} Total
    Males 20,988 9,539 30,527
    Females 13,026 1,669 14,695
    Total 34,014 11,208 45,222
    Adult Dataset (Race) y≤50Ky \le 50\text{K} y>50Ky > 50\text{K} Total
    Amer.-Indian/Eskimo 382 53 435
    Asian/Pacific-Islander 934 369 1,303
    White 28,696 10,207 38,903
    Black 3,694 534 4,228
    Other 308 45 353
    Total 34,014 11,208 45,222
    Bank Dataset (Age) No Yes Total
    25≤age≤6025 \le \text{age} \le 60 35,240 3,970 39,210
    age<25\text{age} < 25 or age>60\text{age} > 60 1,308 670 1,978
    Total 36,548 4,640 41,188

    In the Adult dataset, males have a positive outcome rate (y>50Ky > 50\text{K}) of 31.2%31.2\% compared to 11.4%11.4\% for females. In the Bank dataset, non-protected clients (age <25<25 or >60>60) have a subscription rate of 33.9%33.9\%, whereas the protected demographic (25≤age≤6025 \le \text{age} \le 60) has a subscription rate of only 10.1%10.1\%.

  10. Knowl 10 — Analytical Disconnect Between Decision Boundary Covariance and the $p\%$-Rule

    limitation

    While bounding the decision boundary covariance empirically produces monotonic improvements in the p%p\%-rule (and decreases the Calders-Verwer score), there is no closed-form analytical mapping between a specific covariance threshold cc and the exact resulting p%p\%-rule. The exact relationship depends on the data distribution, margin geometry, and specific classifier chosen.

    Additionally, when protected demographic subgroups are heavily underrepresented in the training data (e.g., the 'Other' racial category in Adult accounting for only 0.8%0.8\% of samples), the empirical covariance estimator exhibits high sample variance, which impairs the classifier's ability to achieve exact fairness for rare subgroups.

Coverage note — None was omitted; all key theoretical formulations, algorithmic SVM/logistic specializations, empirical findings, and limitations were captured.

References

  1. 1.Adult data. http://tinyurl.com/UCI-Adult, 1996.
  2. 2.Bank data. http://tinyurl.com/UCI-Bank, 2014.
  3. 3.S. Barocas and A. D. Selbst. Big Data's Disparate Impact. California Law Review, 2016.
  4. 4.E. Bhandari. Big Data Can Be Used To Violate Civil Rights Laws, and the FTC Agrees. https://www.aclu.org/blog/free-future/big-data-can-be-used-violate-civil-rights-laws-and-ftc-agrees, 2016.
  5. 5.D. Biddle. Adverse Impact and Test Validation: A Practitioner's Guide to Valid and Defensible Employment Testing. Gower, 2005.
  6. 6.T. Calders and S. Verwer. Three Naive Bayes Approaches for Discrimination-Free Classification. Data Mining and Knowledge Discovery, 2010.
  7. 7.Civil Rights Act. Civil Rights Act of 1964, Title VII, Equal Employment Opportunities, 1964.
  8. 8.Y. N. Dauphin, R. Pascanu, C. Gulcehre, K. Cho, S. Ganguli, and Y. Bengio. Identifying and Attacking the Saddle Point Problem in High-dimensional Non-convex Optimization. In NIPS, 2014.
  9. 9.C. Dwork, M. Hardt, T. Pitassi, and O. Reingold. Fairness Through Awareness. In ITCSC, 2012.
  10. 10.M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkatasubramanian. Certifying and removing disparate impact. In KDD, 2015.
  11. 11.G. Goh, A. Cotter, M. Gupta, and M. Friedlander. Satisfying Real-world Goals with Dataset Constraints. In NIPS, 2016.
  12. 12.F. Kamiran and T. Calders. Classifying without Discriminating. In IC4, 2009.
  13. 13.F. Kamiran and T. Calders. Classification with No Discrimination by Preferential Sampling. In BENELEARN, 2010.
  14. 14.T. Kamishima, S. Akaho, H. Asoh, and J. Sakuma. Fairness-aware Classifier with Prejudice Remover Regularizer. In PADM, 2011.
  15. 15.B. T. Luong, S. Ruggieri, and F. Turini. kNN as an Implementation of Situation Testing for Discrimination Discovery and Prevention. In KDD, 2011.
  16. 16.C. Muñoz, M. Smith, and D. Patil. Big Data: A Report on Algorithmic Systems, Opportunity, and Civil Rights. Executive Office of the President. The White House., 2016.
  17. 17.D. Pedreschi, S. Ruggieri, and F. Turini. Discrimination-aware Data Mining. In KDD, 2008.
  18. 18.J. Podesta, P. Pritzker, E. Moniz, J. Holdren, and J. Zients. Big Data: Seizing Opportunities, Preserving Values. Executive Office of the President. The White House., 2014.
  19. 19.Ricci vs. DeStefano. U.S. Supreme Court, 2009.
  20. 20.A. Romei and S. Ruggieri. A Multidisciplinary Survey on Discrimination Analysis. KER, 2014.
  21. 21.B. Schölkopf and A. J. Smola. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. MIT press, 2002.
  22. 22.L. Sweeney. Discrimination in Online Ad Delivery. ACM Queue, 2013.
  23. 23.M. B. Zafar, I. Valera, M. G. Rodriguez, and K. P. Gummadi. Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatment. In WWW, 2017.
  24. 24.R. Zemel, Y. Wu, K. Swersky, T. Pitassi, and C. Dwork. Learning Fair Representations. In ICML, 2013.

Citation

MLA
Zafar, M. B., et al. “Fairness Constraints: Mechanisms for Fair Classification”. arXiv, 2015, http://arxiv.org/abs/1507.05259v5.
APA
Zafar, M. B., Valera, I., Rodriguez, M. G., & Gummadi, K. P. (2015). Fairness Constraints: Mechanisms for Fair Classification. arXiv. http://arxiv.org/abs/1507.05259v5
Chicago
Zafar, M. B., I. Valera, M. G. Rodriguez, and K. P. Gummadi. 2015. “Fairness Constraints: Mechanisms for Fair Classification”. arXiv. http://arxiv.org/abs/1507.05259v5.
Harvard
Zafar, M.B. et al. (2015) “Fairness Constraints: Mechanisms for Fair Classification”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1507.05259v5.
Vancouver
1. Zafar MB, Valera I, Rodriguez MG, Gummadi KP (2015) Fairness Constraints: Mechanisms for Fair Classification. arXiv

BibTeX

@article{zafar2015fairness,
  title = {Fairness Constraints: Mechanisms for Fair Classification},
  author = {Zafar, Muhammad Bilal and Valera, Isabel and Rodriguez, Manuel Gomez and Gummadi, Krishna P.},
  year = {2015},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1507.05259v5},
  eprint = {1507.05259}
}
Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: Authors