Mitigating Bias in Set Selection with Noisy Protected Attributes

Mehrotra, Anay; Celis, L. Elisa

doi:10.1145/3442188.3445887

Cited by 39 publications

(29 citation statements)

References 29 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This effect has been observed multiple times in the literature [22,32,35,39]. Specific solutions have been proposed, e.g., when only the protected attribute itself is corrupted or noisy [7,31,36,49]. However, as shown in [28], full protection against malicious manipulations of the training data is provably impossible when only one dataset is available.…”

Section: Preliminaries and Related Work 21 Fair Classificationmentioning

confidence: 99%

FLEA: Provably Fair Multisource Learning from Unreliable Training Data

Iofinova¹,

Konstantinov²,

Lampert³

2021

Preprint

View full text Add to dashboard Cite

Fairness-aware learning aims at constructing classifiers that not only make accurate predictions, but do not to discriminate against specific groups. It is a fast-growing area of machine learning with far-reaching societal impact. However, existing fair learning methods are vulnerable to accidental or malicious artifacts in the training data, which can cause them to unknowingly produce unfair classifiers. In this work we address the problem of fair learning from unreliable training data in the robust multisource setting, where the available training data comes from multiple sources, a fraction of which might be not representative of the true data distribution. We introduce FLEA, a filtering-based algorithm that allows the learning system to identify and suppress those data sources that would have a negative impact on fairness or accuracy if they were used for training. We show the effectiveness of our approach by a diverse range of experiments on multiple datasets. Additionally we prove formally that -given enough data-FLEA protects the learner against unreliable data as long as the fraction of affected data sources is less than half.* Equal contribution Preprint. Under review.

show abstract

Section: Preliminaries and Related Work 21 Fair Classificationmentioning

confidence: 99%

FLEA: Provably Fair Multisource Learning from Unreliable Training Data

Iofinova¹,

Konstantinov²,

Lampert³

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Most of the recent works utilize average performance indicators broken down by sensitive groups. The metrics themselves measure disparities between groups in terms of predicted positive value (PPV), True/False positive rates (TPR, FPR) [5], error rates [11] or risk difference [57]. There is no consensus nor general guidance for choosing the metric depending on context.…”

Section: Further Context and Related Workmentioning

confidence: 99%

Fairness Indicators for Systematic Assessments of Visual Feature Extractors

Goyal¹,

Soriano²,

Hazırbaş³

et al. 2022

Preprint

View full text Add to dashboard Cite

“…as they might be if labelled using crowdsourced workers) and give conditions (such as conditional independence of the noisy sensitive features and the predicted label) under which post-processing a fixed classifier to equalize false positive or negative rates as measured under the proxy will reduce the true disparity between false positive or negative rates subject to the true sensitive features. In similar noise models, propose robust optimization based approaches to fairness constrained training with noisy sensitive features and Mehrotra and Celis [2021] consider the problem of fair subset selection. Lahoti et al [2020] propose to solve a minimax optimization problem over an enormous set of "computationally identifiable" subgroups, under the premise that if there exists a good proxy for a sensitive feature, then it will be included as one of these computationally identifiable groups defined with respect to the other features.…”

Section: Related Workmentioning

confidence: 99%

Multiaccurate Proxies for Downstream Fairness

Diana¹,

Gill²,

Kearns³

et al. 2021

Preprint

View full text Add to dashboard Cite

We study the problem of training a model that must obey demographic fairness conditions when the sensitive features are not available at training time -in other words, how can we train a model to be fair by race when we don't have data about race? We adopt a fairness pipeline perspective, in which an "upstream" learner that does have access to the sensitive features will learn a proxy model for these features from the other attributes. The goal of the proxy is to allow a general "downstream" learnerwith minimal assumptions on their prediction task -to be able to use the proxy to train a model that is fair with respect to the true sensitive features. We show that obeying multiaccuracy constraints with respect to the downstream model class suffices for this purpose, and provide sample-and oracle efficientalgorithms and generalization bounds for learning such proxies. In general, multiaccuracy can be much easier to satisfy than classification accuracy, and can be satisfied even when the sensitive features are hard to predict.

show abstract

Mitigating Bias in Set Selection with Noisy Protected Attributes

Cited by 39 publications

References 29 publications

FLEA: Provably Fair Multisource Learning from Unreliable Training Data

FLEA: Provably Fair Multisource Learning from Unreliable Training Data

Fairness Indicators for Systematic Assessments of Visual Feature Extractors

Multiaccurate Proxies for Downstream Fairness

Contact Info

Product

Resources

About