Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fairness Without Harm: An Influence-Guided Active Sampling Approach

Jinlong Pang, Jialu Wang, Zhaowei Zhu, Yuanshun Yao, Chen Qian, Yang Liu

TL;DR

This work proposes a tractable active data sampling algorithm that does not rely on training group annotations, instead only requiring group annotations on a small validation set, and provides the upper bound of generalization error and risk disparity as well as the corresponding connections.

Abstract

The pursuit of fairness in machine learning (ML), ensuring that the models do not exhibit biases toward protected demographic groups, typically results in a compromise scenario. This compromise can be explained by a Pareto frontier where given certain resources (e.g., data), reducing the fairness violations often comes at the cost of lowering the model accuracy. In this work, we aim to train models that mitigate group fairness disparity without causing harm to model accuracy. Intuitively, acquiring more data is a natural and promising approach to achieve this goal by reaching a better Pareto frontier of the fairness-accuracy tradeoff. The current data acquisition methods, such as fair active learning approaches, typically require annotating sensitive attributes. However, these sensitive attribute annotations should be protected due to privacy and safety concerns. In this paper, we propose a tractable active data sampling algorithm that does not rely on training group annotations, instead only requiring group annotations on a small validation set. Specifically, the algorithm first scores each new example by its influence on fairness and accuracy evaluated on the validation dataset, and then selects a certain number of examples for training. We theoretically analyze how acquiring more data can improve fairness without causing harm, and validate the possibility of our sampling approach in the context of risk disparity. We also provide the upper bound of generalization error and risk disparity as well as the corresponding connections. Extensive experiments on real-world data demonstrate the effectiveness of our proposed algorithm. Our code is available at https://github.com/UCSC-REAL/FairnessWithoutHarm.

Fairness Without Harm: An Influence-Guided Active Sampling Approach

TL;DR

This work proposes a tractable active data sampling algorithm that does not rely on training group annotations, instead only requiring group annotations on a small validation set, and provides the upper bound of generalization error and risk disparity as well as the corresponding connections.

Abstract

The pursuit of fairness in machine learning (ML), ensuring that the models do not exhibit biases toward protected demographic groups, typically results in a compromise scenario. This compromise can be explained by a Pareto frontier where given certain resources (e.g., data), reducing the fairness violations often comes at the cost of lowering the model accuracy. In this work, we aim to train models that mitigate group fairness disparity without causing harm to model accuracy. Intuitively, acquiring more data is a natural and promising approach to achieve this goal by reaching a better Pareto frontier of the fairness-accuracy tradeoff. The current data acquisition methods, such as fair active learning approaches, typically require annotating sensitive attributes. However, these sensitive attribute annotations should be protected due to privacy and safety concerns. In this paper, we propose a tractable active data sampling algorithm that does not rely on training group annotations, instead only requiring group annotations on a small validation set. Specifically, the algorithm first scores each new example by its influence on fairness and accuracy evaluated on the validation dataset, and then selects a certain number of examples for training. We theoretically analyze how acquiring more data can improve fairness without causing harm, and validate the possibility of our sampling approach in the context of risk disparity. We also provide the upper bound of generalization error and risk disparity as well as the corresponding connections. Extensive experiments on real-world data demonstrate the effectiveness of our proposed algorithm. Our code is available at https://github.com/UCSC-REAL/FairnessWithoutHarm.
Paper Structure (60 sections, 8 theorems, 45 equations, 7 figures, 7 tables, 1 algorithm)

This paper contains 60 sections, 8 theorems, 45 equations, 7 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Fairness-accuracy tradeoff
  4. Active learning
  5. Fair classifiers without demographics
  6. Fair classification
  7. Preliminaries
  8. Problem setup
  9. Fairness definition
  10. Improving fairness without harm via data influential sampling
  11. Finding influential examples
  12. Influence of accuracy component
  13. Influence of fairness component
  14. Intuitions
  15. Training sensitive attributes are not disclosed
  16. ...and 45 more sections

Key Result

Proposition 3.1

(Informal) Under appropriate conditions, the risk disparity can serve as a lower bound for fairness disparities based on common fairness definitions, such as DP and EOd.

Figures (7)

  • Figure 1: We compare the Pareto frontiers between the model trained with scarce data and that trained with rich data. Acquiring more data is capable of shifting the Pareto frontier toward lower disparity and lower error rates. In consequence, we can reach a new trade-off point that offers improved fairness and accuracy simultaneously, surpassing the original trade-off point.
  • Figure 2: Main results on CelebA, Adult and Compas datasets. The Y axis shows fairness_violation; X axis denotes test_accuracy. CelebA: Four binary targets: Smiling, Attractive, Young, and Big_Nose; Sensitive attribute: gender. Adult: Binary target: Income; Sensitive attribute: Age. Compas: Binary target: Recidivism; Sensitive attribute: Race. We select two fairness metrics DP and Eop to measure fairness violations for each setting. The vertical dotted line at the random baseline accuracy helps easily identify which results achieve fairness without sacrificing performance (accuracy).
  • Figure 3: Left: The impact of label budgets on the test accuracy & DP gap in the CelebA dataset. Right: The impact of the validation set size on (test_accuracy, fairness_violation) results.
  • Figure 4: We validate how accurate the first-order estimation of the influence is in comparison to the real influence. The $x$-axis represents the actual influence per sample, and the $y$-axis represents the estimated influence. We observe that while some of the examples are away from the diagonal line (which indicates the estimation is inaccurate), the estimated influences for most of the data samples are very close to their actual influence values.
  • Figure 5: The impact of label budgets on the test accuracy & DP gap in the CelebA dataset. The binary classification targets is Smiling.
  • ...and 2 more figures

Theorems & Definitions (22)

  • Definition 3.1: Risk disparity hashimoto2018fairnesszafar2019fairnessagarwal2019fair
  • Definition 3.2: Demographic Parity (DP)
  • Definition 3.3: Equalized Odds (EOd hardt2016equality
  • Proposition 3.1
  • Remark 3.1: Connections to other fairness definitions
  • Lemma 4.1
  • Lemma 4.2
  • Theorem 5.1: Generalization error bound
  • Theorem 5.2: Upper bound of risk disparity
  • proof
  • ...and 12 more