Learning Smooth and Fair Representations

Xavier Gitiaux, Huzefa Rangwala

Abstract: This paper explores the statistical properties of fair representation learning, a pre-processing method that preemptively removes the correlations between features and sensitive attributes by mapping features to a fair representation space. We show that the demographic parity of a representation can be certified from a finite sample if and only if the mapping guarantees that the chi-squared mutual information between features and representations is finite for distributions of the features. Empirically, we find that smoothing representations with an additive Gaussian white noise provides generalization guarantees of fairness certificates, which improves upon existing fair representation learning approaches.

Introduction

In the field of fair machine learning, data controllers often seek to share data with third-party processors while ensuring that sensitive attributes (e.g., race, gender) cannot be exploited. The standard approach is to learn a "fair representation"—a transformation of the raw data that preserves utility while stripping away sensitive correlations. However, existing adversarial methods for learning these representations often suffer from a lack of generalization: a representation that appears fair to a training adversary may still allow a different downstream processor to discriminate, rendering fairness certificates "unreliable."

Methodology: Smoothing with AGWN

To address the brittleness of adversarial fairness, this research introduces a mathematically grounded smoothing approach. The core theoretical contribution is the $\chi^2$ condition: for a fairness certificate to be reliable across all possible downstream processors, the $\chi^2$ mutual information between the original features ($X$) and the representation ($Z$) must be finite. Standard neural encoders often fail this condition, allowing downstream users to reconstruct sensitive information from finite samples.

The proposed framework for fair representation learning. An encoder $t(X)$ maps features to a latent space $Z$, smoothed by noise $\epsilon$. Downstream processors $h_i(Z)$ and an auditor $a(Z)$ evaluate the representation's utility and fairness.

The authors propose the Additive Gaussian White Noise (AGWN) channel to satisfy this condition. Instead of a deterministic output, the encoder produces $t(X) + \epsilon$, where $\epsilon$ is Gaussian noise. This noise "smoothes" the distribution, allowing the use of a simpler "plug-in" auditor. This auditor uses Monte Carlo sampling to estimate representation density, enabling a differentiable and theoretically robust fairness loss.

Experimental Setup

The approach was evaluated across four datasets:

  • Swiss Roll & DSprites: Synthetic benchmarks where unfairness is spatially or structurally encoded.
  • UCI Adult & Heritage Health: Real-world datasets (census and clinical data) where gender serves as the sensitive attribute.

Key Results and Reliability

The primary metric of success is the reliability of the fairness certificate. A certificate is considered "reliable" if the empirical bound calculated during training remains an upper bound for the actual disparity observed by a downstream processor.

Empirical certificate reliability for AGWN ($\sigma = 0.05$). The empirical certificate acts as a valid upper bound for processor disparity, keeping the model in the 'Reliable' zone.

In experiments on the Swiss Roll and DSprites datasets, AGWN-trained representations consistently stayed within the "Reliable" zone. In contrast, adversarial methods like AdvCE and AdvL1 frequently fell into the "Unreliable" zone, significantly underestimating the potential for downstream discrimination. Furthermore, the authors proved that these smoothed representations achieve a convergence rate of $O(n^{-1/2})$, ensuring that empirical fairness measurements scale predictably with sample size.

Discussion and Limitations

While AGWN provides superior reliability, it introduces a "Fairness-Utility-Reliability" trilemma. The choice of the noise parameter $\sigma$ is critical; insufficient noise leads to unreliable certificates, while excessive noise degrades data utility (reconstruction accuracy). Future work remains to automate the selection of $\sigma$ and to extend these proofs beyond Demographic Parity to other metrics like Equalized Odds. Additionally, scaling AGWN to high-dimensional data like high-resolution video presents a practical challenge for maintaining reconstruction quality.

Summary updated

Abstract

This paper explores the statistical properties of fair representation learning, a pre-processing method that preemptively removes the correlations between features and sensitive attributes by mapping features to a fair representation space. We show that the demographic parity of a representation can be certified from a finite sample if and only if the mapping guarantees that the chi-squared mutual information between features and representations is finite for distributions of the features. Empirically, we find that smoothing representations with an additive Gaussian white noise provides generalization guarantees of fairness certificates, which improves upon existing fair representation learning approaches.
BibTeX
@inproceedings{gitiaux2021learning,
  author = {Xavier Gitiaux and Huzefa Rangwala},
  title = {{Learning Smooth and Fair Representations}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2021},
  url = {https://openreview.net/forum?id=fV726shEdI},
}
Figure