Empirical Risk Minimization under Fairness Constraints

By Michele Donini, Luca Oneto, Shai Ben-David, John Shawe-Taylor, Massimiliano Pontil. 2018

500 citations recorded by the source corpus.

We address the problem of algorithmic fairness: ensuring that sensitive variables do not unfairly influence the outcome of a classifier. We present an approach based on empirical risk minimization, which incorporates a fairness constraint into the learning problem. It encourages the conditional risk of the learned classifier to be approximately constant with respect to the sensitive variable. We derive both risk and fairness bounds that support the statistical consistency of our approach. We specify our approach to kernel methods and observe that the fairness requirement implies an orthogonality constraint which can be easily added to these methods. We further observe that for linear models the constraint translates into a simple data preprocessing step. Experiments indicate that the method is empirically effective and performs favorably against state-of-the-art approaches.

Subfields: Artificial Intelligence, Safety Research, Optimization & Training, AI Ethics, Fairness & Society

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Safety Research

In Optimization & Training

In AI Ethics, Fairness & Society

Citations in the corpus

CERTIFAI: Counterfactual Explanations for Robustness, Transparency, Interpretability, and Fairness of Artificial Intelligence models (2019, background): “Most fairness auditing models focus on single features (e.g., [Hardt et al., 2016; Donini et al., 2018]).”
Wasserstein Fair Classification (2019, methodology): “Other work similar in spirit adds penalties to the objective (Donini et al., 2018; Komiyama et al., 2018).”
Taking Advantage of Multitask Learning for Fair Classification (2018, methodology): “We selected the best hyperparameters9 by the two steps 10-fold cross validation (CV) procedure described in [16].”
Characterizing Fairness Over the Set of Good Models Under Selective Labels (2021, background): “An influential literature on fair classification and fair regression constructs prediction functions that minimize loss subject to a predictive fairness constraint chosen by the decision maker (Dwork et al., 2012; Zemel et al., 2013; Hardt et al., 2016; Menon & Williamson, 2018; Donini et al., 2018;”
A Reductions Approach to Fair Classification (2018, background): “Copyright 2018 by the author(s). methods, often by relaxing the desired deﬁnitions of fairness, and only enforcing weaker constraints, such as lack of correlation (e.g., Woodworth et al., 2017; Zafar et al., 2017; Johnson et al., 2016; Kamishima et al., 2011; Donini et al., 2018).”
Optimization with Non-Differentiable Constraints with Applications to Fairness, Recall, Churn, and Other Goals (2018, methodology): “AUC can be approximated as a set of rate constraints, using the approximation proposed in Eban et al. (2017). Churn: Given a current classifier, the churn of a new classifier on a specific distribution of examples is the probability that the new classifier decision differs from the current classifie”
Fairness in Machine Learning: A Survey (2020, unknown): “[85, 131]), TPR/FPR of protected groups (e.”
Identifying and Correcting Label Bias in Machine Learning (2019, background): “There is also recent work similar in spirit which encourages fairness by adding penalties to the objective; e.g, Donini et al. (2018) studies this for kernel methods and Komiyama et al. (2018) for linear models.”
Fairness risk measures (2019, methodology): “…our contributions are as follows: ( C1 ) building on notions of fairness in terms of subgroup errors [Zafar et al., 2017b, Dwork et al., 2018, Donini et al., 2018], we provide a new deﬁnition of fairness (Deﬁnition 3) requiring the average losses (or risks ) for each sensitive feature…”
Mitigating Political Bias in Language Models Through Reinforced Calibration (2021, background): “Finally, there is related art in machine learning fairness research seeking to produce “fair” classiﬁers or unbiased feature representations (Zhao et al. 2019; Donini et al. 2018; Misra et al. 2016; Kamishima et al. 2012).”
Out-of-Distribution Generalization via Risk Extrapolation (REx) (2020, background): “female) has been proposed as a definition of fairness (Donini et al., 2018), generalizing the equal opportunity definition of fairness (Hardt et al.”
Training Well-Generalizing Classifiers for Fairness Metrics and Other Data-Dependent Constraints (2018, background): “ heuristic for improving constraint generalization. 2 Related Work While several recent papers have proved generalization bounds for constrained problems [e.g. Goh et al., 2016, Agarwal et al., 2018, Donini et al., 2018], the problem of improving constraint generalization is a fairly new one, having”
Tilted Empirical Risk Minimization (2020, background): “Constraints could also be imposed to promote fairness (Baharlouei et al., 2020; Donini et al., 2018; Hardt et al., 2016; Rezaei et al., 2020; Zafar et al., 2017).”
Noise-tolerant fair classification (2019, background): “…Zafar et al., 2017b,a; Kusner et al., 2017; Speicher et al., 2018; Zhang & Bareinboim, 2018; Heidari et al., 2019), and on means of achieving them (Zemel et al., 2013; Zafar et al., 2017b; Calmon et al., 2017; Dwork et al., 2018; Agarwal et al., 2018; Donini et al., 2018; Heidari et al., 2018).”
Robust Optimization for Fairness with Noisy Protected Groups (2020, methodology): “We begin with the training problem for incorporating group-based fairness criteria in a learning setting [27, 32, 17, 2, 14].”
Uncertainty as a Form of Transparency: Measuring, Communicating, and Using Uncertainty (2020, methodology): “In-processing methods modify the learning objective by introducing constraints through which the resulting classifier can achieve the desired fairness properties [1, 43, 192].”