TL;DR
This paper introduces a novel framework for quantifying algorithmic unfairness by adapting inequality indices from economics. It allows for a unified measurement of both individual and group unfairness, revealing critical trade-offs between different fairness notions and accuracy that were previously overlooked.
What Problem Does This Paper Address?
As algorithmic decision-making systems become more prevalent in critical areas like criminal justice and credit scoring, ensuring their fairness is paramount. While existing research has proposed various definitions of fairness, a significant gap exists in how to quantify unfairness. Specifically, it's challenging to answer the question: "Which of two unfair algorithms is more unfair?" Current methods often focus on binary fairness conditions (fair/unfair) rather than providing a nuanced measure. This lack of a principled, generalizable quantification method hinders the comparison of different algorithms and the understanding of trade-offs, particularly between individual and group fairness, and between fairness and predictive accuracy.
Methods
The core innovation of this paper is the application of inequality indices from economics and social welfare to measure algorithmic unfairness. The authors propose defining a "benefit function" that maps algorithmic outcomes to a non-negative real number, representing how desirable an outcome is for an individual. By adapting this benefit function, existing inequality indices (like the Generalized Entropy index) can then be used to quantify how unequally these benefits are distributed across individuals and groups.
Key aspects of their approach include:
- Unified Framework: The method allows for quantifying unfairness at both the individual and group levels using the same underlying principles.
- Subgroup Decomposability: They leverage the property of subgroup decomposability in inequality indices, which allows the overall unfairness measure to be broken down into a "between-group" component (inequality between group averages) and a "within-group" component (inequality within each group). This is crucial for understanding trade-offs.
- Benefit Function Adaptation: The choice of benefit function is tailored to specific fairness notions. For example, for statistical parity, benefits are assigned to reflect acceptance, while for individual fairness, benefits capture the discrepancy between desired and actual outcomes.
Datasets
The authors empirically validate their approach using two well-known real-world datasets:
- Adult Income Dataset: Used for predicting whether an individual earns above or below $50,000 USD. Sensitive features considered include gender and race.
- ProPublica COMPAS Dataset: Used for predicting the likelihood of a criminal defendant re-offending. Sensitive features considered include gender, race, and age.
They employ standard machine learning models like logistic regression, SVM, and random forest, and use cross-validation for hyperparameter tuning and evaluation.
Key Insights and Ideas
- Economics as a Blueprint for Fairness: Inequality indices, with their well-established axiomatic foundations (anonymity, population invariance, transfer principle, zero-normalization), provide a principled and generalizable way to measure algorithmic unfairness.
- The Hidden Cost of Focusing Solely on Group Fairness: Minimizing only the "between-group" unfairness (e.g., ensuring equal false positive rates across racial groups) can paradoxically increase "within-group" unfairness, leading to higher overall individual unfairness. This highlights a critical tension often ignored by existing fairness-aware algorithms.
- Granularity Matters for Group Fairness: The contribution of between-group unfairness to overall unfairness increases with the number of defined groups (e.g., considering race and gender together vs. just race). This suggests that "fairness gerrymandering" can occur if groups are not carefully considered.
- Accuracy vs. Fairness Trade-offs are Nuanced: While an "oracle" classifier (perfectly accurate) is also perfectly fair, imperfect classifiers exhibit complex trade-offs. The optimal threshold for accuracy does not necessarily align with the optimal threshold for fairness, and vice-versa.
- Individual Unfairness Dominates Group Unfairness: In many practical scenarios, the overall individual unfairness is significantly larger than the between-group unfairness, suggesting that focusing solely on group disparities might miss a larger portion of the problem.
Key Results
- The proposed framework successfully quantifies and compares unfairness across different algorithms and fairness notions.
- Empirical results on the Adult and COMPAS datasets demonstrate that existing methods focusing on between-group fairness can lead to increased within-group and overall unfairness.
- For example, in Figure 7, reducing between-group unfairness by tightening constraints led to an increase in within-group unfairness for Whites and, in some cases, an increase in overall unfairness compared to an unconstrained classifier.
- The paper shows that the accuracy-optimal classifier is not always the fairness-optimal classifier, and the accuracy of the fairness-optimal classifier can be arbitrarily worse than the accuracy-optimal classifier in certain scenarios.
Limitations and Future Work
- Benefit Function Specificity: While the framework is general, the specific choice of benefit function for individual fairness (bi = yi - yi + 1) is an illustration. Different contexts might require different benefit assignments.
- Computational Complexity: While not explicitly detailed as a limitation, the application of inequality indices and their decomposition might introduce computational overhead compared to simpler fairness metrics.
- Future Work: The authors suggest extending the framework to unsupervised learning tasks (clustering, ranking) and exploring more complex scenarios with multiple sensitive attributes and intersectional groups. They also highlight the need for further investigation into the practical implications of these trade-offs for algorithm designers.
Summary updated