[Paper Review] Conditional Learning of Fair Representations

Han Zhao, Amanda Coston+2 more|arXiv (Cornell University)|Oct 16, 2019

Ethics and Social Impacts of AI36 references18 citations

TL;DR

This paper proposes a novel algorithm for learning fair representations that simultaneously achieves accuracy parity and equalized odds by leveraging balanced error rate (BER) and conditional alignment of representations. The method optimizes BER on both the target and sensitive attributes, enabling improved utility-fairness trade-offs on balanced datasets without compromising demographic parity.

ABSTRACT

We propose a novel algorithm for learning fair representations that can simultaneously mitigate two notions of disparity among different demographic subgroups in the classification setting. Two key components underpinning the design of our algorithm are balanced error rate and conditional alignment of representations. We show how these two components contribute to ensuring accuracy parity and equalized false-positive and false-negative rates across groups without impacting demographic parity. Furthermore, we also demonstrate both in theory and on two real-world experiments that the proposed algorithm leads to a better utility-fairness trade-off on balanced datasets compared with existing algorithms on learning fair representations for classification.

Motivation & Objective

• To address the open question of whether accuracy parity and equalized odds can coexist in fair representation learning.
• To design a method that ensures fairness across demographic subgroups without sacrificing utility or demographic parity.
• To theoretically and empirically validate that balanced error rate (BER) and conditional alignment jointly enable simultaneous fairness guarantees.
• To demonstrate improved utility-fairness trade-offs on balanced real-world datasets compared to existing methods.

Proposed method

• The algorithm frames fair representation learning as a minimax problem that jointly optimizes BER on both the target variable and the sensitive attribute.
• It enforces conditional alignment of representations across demographic subgroups by aligning the conditional distributions of representations given the target variable.
• The method reduces the minimax problem to cost-sensitive learning, enabling efficient optimization via standard machine learning pipelines.
• Balanced error rate (BER) is used as the primary loss function to ensure equalized error rates across groups and to bound joint error.
• The approach preserves demographic parity by not directly constraining the marginal distribution of the sensitive attribute.
• The algorithm is trained in stages, with each stage solving a linear minimax saddle point problem to iteratively improve fairness and utility.

Experimental results

Research questions

• RQ1 Can accuracy parity and equalized odds be simultaneously achieved in representation learning without compromising demographic parity?
• RQ2 What is the theoretical role of balanced error rate (BER) in ensuring fairness and bounding joint error across groups?
• RQ3 How does conditional alignment of representations contribute to achieving equalized odds and accuracy parity?
• RQ4 Does the proposed method achieve a better utility-fairness trade-off than existing fair representation learning algorithms on balanced datasets?
• RQ5 What is the relationship between BER, equalized odds, and demographic parity in the context of fair representation learning?

Key findings

• The proposed algorithm achieves both accuracy parity and equalized odds simultaneously by combining balanced error rate (BER) and conditional alignment of representations.
• BER is proven to serve as an upper bound on the joint error across demographic subgroups when equalized odds is satisfied.
• The method preserves demographic parity without degradation, meaning demographic parity is maintained 'for free' when equalized odds is achieved.
• On balanced datasets, the algorithm demonstrates a better utility-fairness trade-off compared to existing fair representation learning methods.
• On imbalanced datasets, the algorithm is the only method that achieves accuracy parity, albeit at a cost to overall utility.
• Theoretical analysis confirms that BER plays a fundamental role in ensuring fairness and bounding error, justifying its use over marginal error in the loss function.