Inherent Tradeoffs in Learning Fair Representations

Han Zhao · Geoffrey J. Gordon

78 citations0/0 files storedarXiv (Cornell University)

No readable copy stored yet — this record is metadata only.

No open-access location is known for this record yet.

Literature note Cite in a project Citation map

Ethics and Social Impacts of AI Privacy-Preserving Technologies in Data Explainable Artificial Intelligence (XAI)

Abstract

Real-world applications of machine learning tools in high-stakes domains are often regulated to be fair, in the sense that the predicted target should satisfy some quantitative notion of parity with respect to a protected attribute. However, the exact tradeoff between fairness and accuracy is not entirely clear, even for the basic paradigm of classification problems. In this paper, we characterize an inherent tradeoff between statistical parity and accuracy in the classification setting by providing a lower bound on the sum of group-wise errors of any fair classifiers. Our impossibility theorem could be interpreted as a certain uncertainty principle in fairness: if the base rates differ among groups, then any fair classifier satisfying statistical parity has to incur a large error on at least one of the groups. We further extend this result to give a lower bound on the joint error of any (approximately) fair classifiers, from the perspective of learning fair representations. To show that our lower bound is tight, assuming oracle access to Bayes (potentially unfair) classifiers, we also construct an algorithm that returns a randomized classifier that is both optimal (in terms of accuracy) and fair. Interestingly, when the protected attribute can take more than two values, an extension of this lower bound does not admit an analytic solution. Nevertheless, in this case, we show that the lower bound can be efficiently computed by solving a linear program, which we term as the TV-Barycenter problem, a barycenter problem under the TV-distance. On the upside, we prove that if the group-wise Bayes optimal classifiers are close, then learning fair representations leads to an alternative notion of fairness, known as the accuracy parity, which states that the error rates are close between groups. Finally, we also conduct experiments on real-world datasets to confirm our theoretical findings.

Paper brief

Loading…

Cite this work

Citation unavailable for this work.

BibTeX RIS CSL-JSON

No public copy is stored for this record, so only the metadata is cited.

Citations in this library

Citation map →

Stored only when both papers are in the corpus — 78 citations exist upstream.

References

2018Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification 2012Fairness through awareness 2014Big Data's Disparate Impact 2016Equality of Opportunity in Supervised Learning 2017Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments 2018Mitigating Unwanted Biases with Adversarial Learning 2017Algorithmic Decision Making and the Cost of Fairness 2013Learning Fair Representations

Cited by

2022Quantifying and alleviating political bias in language models 2021Mitigating Political Bias in Language Models through Reinforced Calibration 2022Perturbation Augmentation for Fairer NLP 2021Are Gender-Neutral Queries Really Gender-Neutral? Mitigating Gender Bias in Image Search 2024Implementing equitable and intersectionality‐aware ML in education: A practical guide 2021Removing Spurious Features can Hurt Accuracy and Affect Groups Disproportionately

Explore the map →

Papers that cite the same references — the relational fallback while embeddings are still backfilling.

Imported from OpenAlex (CC0) — nothing here is generated.