[Paper Review] Understanding Instance-Level Impact of Fairness Constraints

Jialu Wang, Xin Wang+1 more|arXiv (Cornell University)|Jun 30, 2022

Gender Politics and Representation4 citations

TL;DR

This paper introduces the fairness influence function, a method to quantify the instance-level impact of training examples under fairness constraints by leveraging influence functions. It demonstrates that pruning high-influence examples based on fairness scores reduces fairness violations with minimal accuracy drop, achieving a favorable trade-off on tabular, image, and NLP datasets.

ABSTRACT

A variety of fairness constraints have been proposed in the literature to mitigate group-level statistical bias. Their impacts have been largely evaluated for different groups of populations corresponding to a set of sensitive attributes, such as race or gender. Nonetheless, the community has not observed sufficient explorations for how imposing fairness constraints fare at an instance level. Building on the concept of influence function, a measure that characterizes the impact of a training example on the target model and its predictive performance, this work studies the influence of training examples when fairness constraints are imposed. We find out that under certain assumptions, the influence function with respect to fairness constraints can be decomposed into a kernelized combination of training examples. One promising application of the proposed fairness influence function is to identify suspicious training examples that may cause model discrimination by ranking their influence scores. We demonstrate with extensive experiments that training on a subset of weighty data examples leads to lower fairness violations with a trade-off of accuracy.

Motivation & Objective

• To understand how individual training instances influence model fairness when fairness constraints are applied.
• To develop a method for identifying and auditing high-impact training examples that contribute to model discrimination.
• To enable data preprocessing strategies that prioritize instances with high fairness-influence for bias mitigation.
• To evaluate the trade-off between fairness and accuracy when pruning training data based on fairness-influence scores.

Proposed method

• Adapt the influence function framework to measure the change in model predictions due to removing a single training example under fairness constraints.
• Formulate fairness-influence as a kernelized combination of training examples under specific assumptions, enabling efficient computation.
• Apply first-order approximation of influence functions to estimate the impact of each training example on fairness and accuracy.
• Rank training instances by their fairness-influence scores and prune low-influence examples to reduce fairness violations.
• Implement the method across diverse domains: tabular data (Adult), images (CelebA), and NLP (Jigsaw Toxicity dataset).
• Use standard models (MLP, ResNet-18, BERT) with fairness constraints and evaluate performance using accuracy and fairness violation metrics.

Experimental results

Research questions

• RQ1 How do individual training instances influence model fairness when fairness constraints are enforced?
• RQ2 Can fairness-influence scores identify training examples that disproportionately contribute to unfair predictions?
• RQ3 What is the trade-off between fairness and accuracy when pruning training data based on fairness-influence?
• RQ4 How does the performance of fairness-influenced pruning compare to accuracy-based or random pruning across different data modalities?
• RQ5 Can fairness-influence analysis guide preprocessing strategies to reduce bias with minimal accuracy degradation?

Key findings

• Pruning training examples based on fairness-influence scores reduces fairness violations by a large margin, especially when data size is below 40%.
• On the Adult dataset, prune-by-fairness maintains similar accuracy to prune-by-accuracy when data size exceeds 20%, while reducing fairness violations significantly.
• On CelebA, fairness-influenced pruning maintains comparable accuracy trends to accuracy-based pruning but reduces fairness violations more effectively as data size decreases.
• On Jigsaw Toxicity, prune-by-fairness achieves performance similar to the baseline method (lfatr) with lower standard deviation, indicating more stable fairness mitigation.
• Random pruning leads to high fairness violations and poor performance when data size drops below 20%, highlighting the importance of informed data selection.
• The method reveals that extremely small training sets (e.g., <10%) degrade performance significantly, suggesting a lower bound for effective fairness-influenced pruning.