Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Fairer Representations with FairVIC

Charmaine Barker, Daniel Bethell, Dimitar Kazakov

TL;DR

FairVIC tackles bias in deep learning by embedding fairness directly into training through three terms in the loss: variance, invariance, and covariance. By recasting VICReg concepts for supervised bias mitigation and optimizing a constrained total loss, FairVIC achieves consistent improvements across multiple group and individual fairness metrics with minimal accuracy losses. The approach is validated on seven diverse datasets spanning tabular, text, and image modalities, and supported by comprehensive ablations that guide default hyperparameters. The work demonstrates robust, generalisable bias reduction without prescribing a single fairness definition, enabling post hoc evaluation across metrics while maintaining competitive performance.

Abstract

Mitigating bias in automated decision-making systems, particularly in deep learning models, is a critical challenge due to nuanced definitions of fairness, dataset-specific biases, and the inherent trade-off between fairness and accuracy. To address these issues, we introduce FairVIC, an innovative approach that enhances fairness in neural networks by integrating variance, invariance, and covariance terms into the loss function during training. Unlike methods that rely on predefined fairness criteria, FairVIC abstracts fairness concepts to minimise dependency on protected characteristics. We evaluate FairVIC against comparable bias mitigation techniques on benchmark datasets, considering both group and individual fairness, and conduct an ablation study on the accuracy-fairness trade-off. FairVIC demonstrates significant improvements ($\approx70\%$) in fairness across all tested metrics without compromising accuracy, thus offering a robust, generalisable solution for fair deep learning across diverse tasks and datasets.

Learning Fairer Representations with FairVIC

TL;DR

FairVIC tackles bias in deep learning by embedding fairness directly into training through three terms in the loss: variance, invariance, and covariance. By recasting VICReg concepts for supervised bias mitigation and optimizing a constrained total loss, FairVIC achieves consistent improvements across multiple group and individual fairness metrics with minimal accuracy losses. The approach is validated on seven diverse datasets spanning tabular, text, and image modalities, and supported by comprehensive ablations that guide default hyperparameters. The work demonstrates robust, generalisable bias reduction without prescribing a single fairness definition, enabling post hoc evaluation across metrics while maintaining competitive performance.

Abstract

Mitigating bias in automated decision-making systems, particularly in deep learning models, is a critical challenge due to nuanced definitions of fairness, dataset-specific biases, and the inherent trade-off between fairness and accuracy. To address these issues, we introduce FairVIC, an innovative approach that enhances fairness in neural networks by integrating variance, invariance, and covariance terms into the loss function during training. Unlike methods that rely on predefined fairness criteria, FairVIC abstracts fairness concepts to minimise dependency on protected characteristics. We evaluate FairVIC against comparable bias mitigation techniques on benchmark datasets, considering both group and individual fairness, and conduct an ablation study on the accuracy-fairness trade-off. FairVIC demonstrates significant improvements () in fairness across all tested metrics without compromising accuracy, thus offering a robust, generalisable solution for fair deep learning across diverse tasks and datasets.
Paper Structure (31 sections, 1 theorem, 7 equations, 7 figures, 8 tables, 1 algorithm)

This paper contains 31 sections, 1 theorem, 7 equations, 7 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. VICReg
  5. Group Fairness Metrics
  6. Individual Fairness
  7. Approach
  8. FairVIC Training
  9. Experiments
  10. Datasets
  11. Comparable Techniques
  12. Evaluation
  13. Core Results Analysis
  14. Individual Fairness Analysis
  15. Language and Image Dataset Results
  16. ...and 16 more sections

Key Result

Theorem 1

Each individual term in FairVIC $L_{var}, L_{inv}, L_{cov}$ is sub-differentiable everywhere in the model's parameters $\theta$.

Figures (7)

  • Figure 1: A qualitative analysis of the absolute differences from the ideal value (e.g., perfect accuracy and fairness) in performance (left) and fairness (right) metrics of comparable techniques on the Adult Income dataset.
  • Figure 2: Network architecture for tabular data, with FairVIC loss components applied at relevant stages.
  • Figure 3: A qualitative analysis of the absolute differences from the ideal value (e.g., perfect accuracy and fairness) in performance (left) and fairness (right) metrics of comparable techniques on the COMPAS and German Credit datasets.
  • Figure 4: Mean feature importances derived from SHAP values for the baseline and FairVIC models across the three tabular datasets. The protected attribute (green) and its strong proxy variables (black) are annotated with their exact feature importance values.
  • Figure 5: An example latent space visualization from one random seed of a baseline model and a FairVIC model on the Adult Income dataset. Subgroup (1) represents male individuals, and subgroup (0) represents female individuals.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Theorem 1
  • proof