Evaluating Fairness of Machine Learning Models Under Uncertain and Incomplete Information

  title={Evaluating Fairness of Machine Learning Models Under Uncertain and Incomplete Information},
  author={Pranjal Awasthi and Alex Beutel and Matthaeus Kleindessner and Jamie H. Morgenstern and Xuezhi Wang},
  journal={Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency},
Training and evaluation of fair classifiers is a challenging problem. This is partly due to the fact that most fairness metrics of interest depend on both the sensitive attribute information and label information of the data points. In many scenarios it is not possible to collect large datasets with such information. An alternate approach that is commonly used is to separately train an attribute classifier on data with sensitive attribute information, and then use it later in the ML pipeline to… 

Figures and Tables from this paper

Estimating and Controlling for Fairness via Sensitive Attribute Predictors

This work demonstrates that in settings where sensitive attributes are unknown, one can still reliably estimate and ultimately control for fairness by using proxy sensitive attributes derived from a sensitive attribute predictor.

Mitigating Algorithmic Bias with Limited Annotations

According to the evaluation on benchmark datasets, APOD outperforms the state-of-the-arts baseline methods under the limited annotation budget, and shows comparable performance to fully annotated bias mitigation, which demonstrates that APOD could benefit real-world applications when sensitive information is limited.

Measuring Model Fairness under Noisy Covariates: A Theoretical Perspective

This work studies using a proxy for the covariate variable and presents a theoretical analysis that aims to characterize weaker conditions under which accurate fairness evaluation is possible and expands the understanding of scenarios where measuring model fairness via proxies can be an effective approach.

Measuring Fairness under Unawareness of Sensitive Attributes: A Quantification-Based Approach

This work tackles the problem of measuring group fairness under unawareness of sensitive attributes, by using techniques from quantification, a supervised learning task concerned with directly providing group-level prevalence estimates (rather than individual-level class labels), and shows that quantification approaches are particularly suited to tackle the fairness-under-unawareness problem.

Measuring Fairness under Unawareness via Quantification

This work tackles the problem of measuring group fairness under unawareness of sensitive attributes, by using techniques from quantification, a supervised learning task concerned with directly providing group-level prevalence estimates (rather than individual-level class labels).

Measuring Fairness of Rankings under Noisy Sensitive Information

This work investigates the problem of measuring group fairness in ranking for a suite of divergence-based metrics in the presence of proxy labels and shows that under certain assumptions, fairness of a ranking can be measured from the proxy labels.

Algorithmic Fairness Datasets: the Story so Far

This work surveys over two hundred datasets employed in algorithmic fairness research, and produces standardized and searchable documentation for each of them, rigorously identifying the three most popular fairness datasets, namely Adult, COMPAS, and German Credit, for which this unifying documentation effort supports multiple contributions.

What-Is and How-To for Fairness in Machine Learning: A Survey, Reflection, and Perspective

It is demonstrated the importance of matching the mission and the means of different types of fairness inquiries on the data generating process, on the predicted outcome, and on the induced impact, respectively, to achieve the intended purpose.

Can Less be More? When Increasing-to-Balancing Label Noise Rates Considered Beneficial

This paper quantifies the trade-offs introduced by increasing a certain group of instances’ label noise rate w.r.t. the loss of label informativeness and the lowered learning difficulties and proposes a detection method that informs us which group of labels might suffer from higher noise without using ground truth labels.

Distributionally Robust Data Join

This work introduces the problem of building a predictor which minimizes the maximum loss over all probability distributions over the original features, auxiliary features, and binary labels, whose Wasserstein distance is r1 away from the empirical distribution over the labeled dataset and r2 away from that of the unlabeled dataset.



Noise-tolerant fair classification

If one measures fairness using the mean-difference score, and sensitive features are subject to noise from the mutually contaminated learning model, then owing to a simple identity the authors only need to change the desired fairness-tolerance, and the requisite tolerance can be estimated by leveraging existing noise-rate estimators from the label noise literature.

Transfer of Machine Learning Fairness across Domains

This work offers new theoretical guarantees of improving fairness across domains, and offers a modeling approach to transfer to data-sparse target domains and gives empirical results validating the theory and showing that these modeling approaches can improve fairness metrics with less data.

Fair Transfer Learning with Missing Protected Attributes

This paper proposes two new weighting methods: prevalence-constrained covariate shift (PCCS) which does not require protected attributes in the target domain and target-fair covariates shift (TFCS), which doesNot require protected Attributes in the source domain and empirically demonstrates their efficacy in two applications.

Equalized odds postprocessing under imperfect group information

This paper investigates to what extent fairness interventions can be effective even when only imperfect information about the protected attribute is available, and identifies conditions on the perturbation that guarantee that the bias of a classifier is reduced even by running equalized odds with the perturbed attribute.

Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatment

A new notion of unfairness, disparate mistreatment, is introduced, defined in terms of misclassification rates, which is proposed for decision boundary-based classifiers and can be easily incorporated into their formulation as convex-concave constraints.

On Fairness and Calibration

It is shown that calibration is compatible only with a single error constraint, and that any algorithm that satisfies this relaxation is no better than randomizing a percentage of predictions for an existing classifier.

Fairness Under Unawareness: Assessing Disparity When Protected Class Is Unobserved

This paper decomposes the biases in estimating outcome disparity via threshold-based imputation into multiple interpretable bias sources, allowing us to explain when over- or underestimation occurs and proposes an alternative weighted estimator that uses soft classification.

Equality of Opportunity in Supervised Learning

This work proposes a criterion for discrimination against a specified sensitive attribute in supervised learning, where the goal is to predict some target based on available features and shows how to optimally adjust any learned predictor so as to remove discrimination according to this definition.

Online Learning with an Unknown Fairness Metric

An algorithm in the adversarial context setting that has a number of fairness violations that depends only logarithmically on $T$, while obtaining an optimal $O(\sqrt{T})$ regret bound to the best fair policy is proposed.

Assessing algorithmic fairness with unobserved protected class using data combination

This paper studies a fundamental challenge to assessing disparate impacts, or performance disparities in general, in practice: protected class membership is often not observed in the data, particularly in lending and healthcare, and provides optimization-based algorithms for computing and visualizing sets of simultaneously achievable pairwise disparties for assessing disparities.