Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AIM: Attributing, Interpreting, Mitigating Data Unfairness

Zhining Liu, Ruizhong Qiu, Zhichen Zeng, Yada Zhu, Hendrik Hamann, Hanghang Tong

TL;DR

The paper addresses the problem of data-level unfairness by introducing AIM, a framework for attributing, interpreting, and mitigating biases encoded in training data. It defines a credibility-aware sample bias criterion and a similarity mechanism based on a comparability graph and random walk to enable per-sample bias attribution with transparent explanations. The approach yields two minimal-edit mitigation strategies, AIM_REM and AIM_AUG, to reduce both group and individual unfairness with minimal predictive utility loss. Extensive experiments on four real-world datasets demonstrate AIM's effectiveness in explaining discrimination, identifying discriminatory samples, and achieving favorable fairness-utility trade-offs. The work advances auditable data-level fairness and offers practical avenues for improving FairML pipelines, with available code for reproducibility.

Abstract

Data collected in the real world often encapsulates historical discrimination against disadvantaged groups and individuals. Existing fair machine learning (FairML) research has predominantly focused on mitigating discriminative bias in the model prediction, with far less effort dedicated towards exploring how to trace biases present in the data, despite its importance for the transparency and interpretability of FairML. To fill this gap, we investigate a novel research problem: discovering samples that reflect biases/prejudices from the training data. Grounding on the existing fairness notions, we lay out a sample bias criterion and propose practical algorithms for measuring and countering sample bias. The derived bias score provides intuitive sample-level attribution and explanation of historical bias in data. On this basis, we further design two FairML strategies via sample-bias-informed minimal data editing. They can mitigate both group and individual unfairness at the cost of minimal or zero predictive utility loss. Extensive experiments and analyses on multiple real-world datasets demonstrate the effectiveness of our methods in explaining and mitigating unfairness. Code is available at https://github.com/ZhiningLiu1998/AIM.

AIM: Attributing, Interpreting, Mitigating Data Unfairness

TL;DR

The paper addresses the problem of data-level unfairness by introducing AIM, a framework for attributing, interpreting, and mitigating biases encoded in training data. It defines a credibility-aware sample bias criterion and a similarity mechanism based on a comparability graph and random walk to enable per-sample bias attribution with transparent explanations. The approach yields two minimal-edit mitigation strategies, AIM_REM and AIM_AUG, to reduce both group and individual unfairness with minimal predictive utility loss. Extensive experiments on four real-world datasets demonstrate AIM's effectiveness in explaining discrimination, identifying discriminatory samples, and achieving favorable fairness-utility trade-offs. The work advances auditable data-level fairness and offers practical avenues for improving FairML pipelines, with available code for reproducibility.

Abstract

Data collected in the real world often encapsulates historical discrimination against disadvantaged groups and individuals. Existing fair machine learning (FairML) research has predominantly focused on mitigating discriminative bias in the model prediction, with far less effort dedicated towards exploring how to trace biases present in the data, despite its importance for the transparency and interpretability of FairML. To fill this gap, we investigate a novel research problem: discovering samples that reflect biases/prejudices from the training data. Grounding on the existing fairness notions, we lay out a sample bias criterion and propose practical algorithms for measuring and countering sample bias. The derived bias score provides intuitive sample-level attribution and explanation of historical bias in data. On this basis, we further design two FairML strategies via sample-bias-informed minimal data editing. They can mitigate both group and individual unfairness at the cost of minimal or zero predictive utility loss. Extensive experiments and analyses on multiple real-world datasets demonstrate the effectiveness of our methods in explaining and mitigating unfairness. Code is available at https://github.com/ZhiningLiu1998/AIM.
Paper Structure (39 sections, 8 equations, 6 figures, 3 tables, 1 algorithm)

This paper contains 39 sections, 8 equations, 6 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Notations.
  4. Methodology
  5. Sample Bias Criterion
  6. Connection to Existing Fairness Notions.
  7. Unfairness Attribution and Interpretation
  8. Computing Self-Explanatory AIM Bias Score.
  9. Interpreting the AIM Bias Score.
  10. Practical Similarity Computation.
  11. AIM for Unfairness Mitigation
  12. AIM$_\texttt{REM}$: Unfairness Removal.
  13. AIM$_\texttt{AUG}$: Fairness Augmentation.
  14. Experiments and Analysis
  15. Datasets.
  16. ...and 24 more sections

Figures (6)

  • Figure 1: Concept and applications of the proposed AIM (Bias Attribution, Interpretation, Mitigation) framework.
  • Figure 2: Compare AIM$_\texttt{REM}$ and AIM$_\texttt{AUG}$ with group fairness baselines. We show the trade-off between utility (x-axis) and unfairness metrics (y-axis) on 4 real-world FairML tasks. Results close to the upper-left corner have better trade-offs, i.e., with low unfairness (x-axis) and high utility (y-axis). Each column corresponds to a FairML task, and each row corresponds to a utility-unfairness metric pair. As AIM's utility-unfairness trade-off can be controlled by the sample removal/augmentation budget, we show its performance with line plots. We show error bars for both utility and unfairness metrics.
  • Figure 3: Evaluation of the AIM bias attribution quality. Removing high-bias samples identified by AIM from the data greatly reduces the discrimination in the model prediction.
  • Figure 4: Synthetic bias detection results. AIM (4th column) can accurately detect ground-truth biased samples (3rd columns) under both group- and individual-level unfairness.
  • Figure 5: AIM bias score reflect dataset unfairness level. Each dot denotes a combination of datasets, sensitive attributes, and train/test split. We report average training sample bias (x-axis) and test unfairness of the model predictions (y-axis).
  • ...and 1 more figures

Theorems & Definitions (4)

  • Definition 3.1: Sample Bias
  • Definition 3.2: Sample Credibility
  • Remark 3.3: Intuition and Example
  • Definition 3.4: Sample Comparability