[Paper Review] Fairness Improvement with Multiple Protected Attributes: How Far Are We?

Zhenpeng Chen, Jie M. Zhang+2 more|arXiv (Cornell University)|Jul 25, 2023

Qualitative Comparative Analysis Research4 citations

TL;DR

This paper evaluates 11 state-of-the-art fairness improvement methods for machine learning across multiple protected attributes, revealing that optimizing fairness for one attribute often degrades fairness for others—impacting up to 88.3% of scenarios—while performance (accuracy) remains largely stable. The study highlights the critical need to assess fairness across intersectional attributes and move beyond single-attribute fairness metrics.

ABSTRACT

Existing research mostly improves the fairness of Machine Learning (ML) software regarding a single protected attribute at a time, but this is unrealistic given that many users have multiple protected attributes. This paper conducts an extensive study of fairness improvement regarding multiple protected attributes, covering 11 state-of-the-art fairness improvement methods. We analyze the effectiveness of these methods with different datasets, metrics, and ML models when considering multiple protected attributes. The results reveal that improving fairness for a single protected attribute can largely decrease fairness regarding unconsidered protected attributes. This decrease is observed in up to 88.3% of scenarios (57.5% on average). More surprisingly, we find little difference in accuracy loss when considering single and multiple protected attributes, indicating that accuracy can be maintained in the multiple-attribute paradigm. However, the effect on F1-score when handling two protected attributes is about twice that of a single attribute. This has important implications for future fairness research: reporting only accuracy as the ML performance metric, which is currently common in the literature, is inadequate.

Motivation & Objective

• To investigate the impact of fairness improvement methods designed for a single protected attribute when applied to multiple protected attributes simultaneously.
• To assess whether fairness improvements for one attribute lead to unintended fairness degradation in unconsidered attributes.
• To evaluate the fairness-performance trade-off when multiple protected attributes are considered, especially in terms of precision and recall.
• To benchmark 11 state-of-the-art fairness methods across diverse datasets, models, and fairness metrics to understand their effectiveness in intersectional fairness settings.

Proposed method

• The study evaluates 11 fairness improvement methods—spanning pre-, in-, and post-processing techniques—selected from both ML and SE literature.
• It applies these methods on five widely used datasets from financial, social, and medical domains, using diverse ML models including logistic regression, random forest, and neural networks.
• The evaluation uses 15 fairness-performance measurements, including multiple group fairness metrics (e.g., statistical parity difference, equal opportunity difference) and performance metrics (accuracy, precision, recall).
• The analysis focuses on intersectional fairness by measuring disparities across subgroups formed by combinations of protected attributes.
• The researchers systematically compare fairness outcomes across all protected attributes, identifying cases where fairness for one attribute deteriorates when optimizing for another.
• They quantify the fairness-performance trade-off by measuring changes in accuracy, precision, and recall when multiple attributes are considered.

Experimental results

Research questions

• RQ1 To what extent does improving fairness for a single protected attribute negatively affect fairness for unconsidered protected attributes when multiple attributes are present?
• RQ2 How do fairness improvement methods perform in terms of intersectional fairness when multiple protected attributes are considered simultaneously?
• RQ3 What is the impact on model performance (accuracy, precision, recall) when fairness is optimized across multiple protected attributes compared to a single attribute?
• RQ4 How do different fairness metrics and model architectures influence the fairness-performance trade-off in multi-attribute settings?

Key findings

• Improving fairness for a single protected attribute leads to a significant decline in fairness for unconsidered attributes in 57.5% of scenarios on average, and up to 88.3% of scenarios across all experiments.
• The degradation in fairness is particularly pronounced for precision and recall, which are affected approximately five and eight times more than accuracy when multiple attributes are considered.
• Despite the substantial fairness trade-offs, accuracy loss remains minimal when transitioning from single- to multiple-protected-attribute fairness improvement, indicating that performance can be preserved.
• The study reveals that current fairness evaluation practices—relying heavily on accuracy alone—are inadequate, as they fail to capture the disproportionate impact on precision and recall in multi-attribute settings.
• There is a strong correlation between protected attributes that may underlie the observed fairness degradation, suggesting that attribute interactions play a key role in fairness trade-offs.
• The benchmarking of 11 methods shows that no single method consistently outperforms others across all datasets and metrics, underscoring the need for context-aware fairness selection.