Counterpart Fairness - Addressing Systematic between-group Differences in Fairness Evaluation

Yifei Wang, Zhengyang Zhou, Liqin Wang, John Laurentiev, Peter Hou, Li Zhou, Pengyu Hong

Abstract

When using machine learning (ML) to aid decision-making, it is critical to ensure that an algorithmic decision is fair, i.e., it does not discriminate against specific individuals/groups, particularly those from underprivileged populations. Existing group fairness methods require equal group-wise measures, which however fails to consider systematic between-group differences. The confounding factors, which are non-sensitive variables but manifest systematic differences, can significantly affect fairness evaluation. To tackle this problem, we believe that a fairness measurement should be based on the comparison between counterparts (i.e., individuals who are similar to each other with respect to the task of interest) from different groups, whose group identities cannot be distinguished algorithmically by exploring confounding factors. We have developed a propensity-score-based method for identifying counterparts, which prevents fairness evaluation from comparing "oranges" with "apples". In addition, we propose a counterpart-based statistical fairness index, termed Counterpart-Fairness (CFair), to assess fairness of ML models. Various empirical studies were conducted to validate the effectiveness of CFair. We publish our code at https://github.com/zhengyjo/CFair.

Introduction

In machine learning, "Group Fairness" typically involves ensuring equal outcomes across protected groups (e.g., race or gender). However, standard metrics like Demographic Parity (DP) often fail to account for systematic differences in background characteristics between these groups. These confounding factors—non-sensitive variables that manifest differently across populations—can lead to misleading fairness evaluations.

Figure 1: Comparison of group distributions and the impact on standard fairness metrics. (A) shows identical distributions, (B) shows substantial systematic differences, and (C) illustrates how biased sampling can distort the observed mean difference.

As illustrated in Figure 1, a simple difference in means may reflect inherent data distributions rather than actual algorithmic bias. Conversely, a model might appear fair on average while being discriminatory toward individuals who are actually similar. This "dilution effect" occurs when the unfair treatment of highly similar individuals is washed out by the noise of the broader, incomparable population.

Methodology: The CFair Framework

To address the "apples-to-oranges" comparison problem, we propose Counterpart Fairness (CFair). This framework identifies "counterparts"—pairs of individuals from different groups who are otherwise identical—to reveal hidden algorithmic biases.

Figure 2: The CFair pipeline: (A) identifying potential confounding factors, (B) using Propensity Score Matching (PSM) to mitigate confounding bias, and (C) performing 1-1 counterpart matching for fairness evaluation.

The CFair pipeline involves three key steps:

  1. Propensity Score Matching (PSM): We train a model to predict an individual's protected attribute based on non-protected features. If the model cannot distinguish which group an individual belongs to (i.e., they have similar propensity scores), they are considered candidates for matching.
  2. 1-1 Counterpart Matching: We use metric learning (Mahalanobis distance) to find the closest possible match in the other group based on baseline characteristics, ensuring truly comparable pairs.
  3. The CFair Index: Once counterparts are identified, we apply a paired t-test to the model's predictions. This allows us to determine if the model is statistically significantly biased against similar individuals, providing a p-value for rigorous assessment.

Empirical Results

We evaluated CFair across four benchmarks: MIMIC-IV (Healthcare), COMPAS (Criminal Justice), Adult Income (Socioeconomic), and German Banking (Finance).

  • Revealing Hidden Unfairness: In the MIMIC and COMPAS datasets, CFair revealed that models were significantly more biased than standard metrics suggested. For COMPAS, the standard DP gap was 0.275, but the Counterpart DP (CDP) gap jumped to 0.442 (p < 0.001).
  • Data Incomparability: In the Adult and German Banking datasets, we found that groups were so systematically different that propensity scores had almost no overlap. This suggests that in some domains, traditional fairness evaluations are fundamentally flawed because the groups are essentially incomparable.
  • Feature Alignment: After matching, the statistical differences between individual features (e.g., age, blood pressure) across groups became non-significant, proving the matching successfully created "apples-to-apples" comparisons.

Conclusion and Limitations

CFair provides a rigorous way to distinguish between disparities caused by data distributions and genuine algorithmic discrimination. However, its reliance on finding enough counterparts means it may be less effective in datasets with extreme systematic differences. Future work will explore one-to-many matching to improve robustness in datasets with limited group overlap.

Summary updated

Abstract

When using machine learning (ML) to aid decision-making, it is critical to ensure that an algorithmic decision is fair, i.e., it does not discriminate against specific individuals/groups, particularly those from underprivileged populations. Existing group fairness methods require equal group-wise measures, which however fails to consider systematic between-group differences. The confounding factors, which are non-sensitive variables but manifest systematic differences, can significantly affect fairness evaluation. To tackle this problem, we believe that a fairness measurement should be based on the comparison between counterparts (i.e., individuals who are similar to each other with respect to the task of interest) from different groups, whose group identities cannot be distinguished algorithmically by exploring confounding factors. We have developed a propensity-score-based method for identifying counterparts, which prevents fairness evaluation from comparing "oranges" with "apples". In addition, we propose a counterpart-based statistical fairness index, termed Counterpart-Fairness (CFair), to assess fairness of ML models. Various empirical studies were conducted to validate the effectiveness of CFair. We publish our code at \url{https://github.com/zhengyjo/CFair}.
BibTeX
@misc{wang2023counterpart,
  author = {Yifei Wang and Zhengyang Zhou and Liqin Wang and John Laurentiev and Peter Hou and Li Zhou and Pengyu Hong},
  title = {{Counterpart Fairness - Addressing Systematic between-group Differences in Fairness Evaluation}},
  year = {2023},
  eprint = {2305.18160},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.18160},
}
Figure