Fairness Beyond Disparate Treatment & Disparate Impact: Learning Classification without Disparate Mistreatment (1610.08452v2)

Abstract: Automated data-driven decision making systems are increasingly being used to assist, or even replace humans in many settings. These systems function by learning from historical decisions, often taken by humans. In order to maximize the utility of these systems (or, classifiers), their training involves minimizing the errors (or, misclassifications) over the given historical data. However, it is quite possible that the optimally trained classifier makes decisions for people belonging to different social groups with different misclassification rates (e.g., misclassification rates for females are higher than for males), thereby placing these groups at an unfair disadvantage. To account for and avoid such unfairness, in this paper, we introduce a new notion of unfairness, disparate mistreatment, which is defined in terms of misclassification rates. We then propose intuitive measures of disparate mistreatment for decision boundary-based classifiers, which can be easily incorporated into their formulation as convex-concave constraints. Experiments on synthetic as well as real world datasets show that our methodology is effective at avoiding disparate mistreatment, often at a small cost in terms of accuracy.

Disparate Mistreatment: A New Approach to Fairness in Classification

The paper "Fairness Beyond Disparate Treatment and Disparate Impact: Learning Classification without Disparate Mistreatment" by Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi introduces a novel concept in the field of fairness-aware machine learning—disparate mistreatment. This notion accounts for fairness in terms of misclassification rates among different social groups. The authors propose intuitive measures to quantify this form of unfairness and incorporate these measures into a flexible framework that can be applied to decision boundary-based classifiers.

Overview of the Research Problem

In automated decision-making systems, historical data often reflects biases inherent in human decision-making. When such biased data is used to train classifiers, the resulting model may perpetuate or even amplify these biases. Existing fairness notions, such as disparate treatment (where decisions vary based on sensitive attributes) and disparate impact (where outcomes disproportionately affect certain groups), have been extensively studied. However, these notions fall short in scenarios where ground truth data is available, and historical decisions can be verified. In such cases, ensuring fairness in terms of equal misclassification rates for different groups becomes essential.

The Concept of Disparate Mistreatment

The authors introduce disparate mistreatment as a measure of fairness defined by the misclassification rates across different groups. Specifically, they propose the following measures to capture disparate mistreatment:

• Overall Misclassification Rate (OMR)
• False Positive Rate (FPR)
• False Negative Rate (FNR)
• False Omission Rate (FOR)
• False Discovery Rate (FDR)

These measures ensure that classifiers do not unfairly disadvantage any group based on their sensitive attributes by having significantly different error rates for these groups.

Methodology

The proposed method integrates fairness constraints into decision boundary-based classifiers, such as logistic regression and support vector machines (SVMs). The key technique involves incorporating covariance constraints to control the degree of disparate mistreatment:

1. Formulation of the Problem: The objective is to minimize the classification loss subject to constraints that ensure equitable misclassification rates across groups. These constraints are captured as a covariance between the sensitive attribute and a misclassification-indicating function.
2. Convex-Concave Programming: Although the fairness constraints are non-convex, the authors reformulate the problem into a Disciplined Convex-Concave Program (DCCP). This approach leverages recent advancements in convex-concave programming to efficiently solve the optimization problem.

Experimental Results

The authors conduct extensive experiments on synthetic and real-world datasets, including the ProPublica COMPAS dataset, known for its reported biases in criminal risk assessments. Key findings include:

• The proposed methodology effectively reduces disparate mistreatment at a modest cost in accuracy.
• The method can simultaneously address multiple fairness criteria (disparate treatment and disparate mistreatment).
• When compared to existing methods, such as the one proposed by Hardt et al., the approach does not use sensitive attributes at decision time, thereby also mitigating disparate treatment.

Implications and Future Work

This research opens new avenues for designing fair classifiers that can account for misclassification biases across different groups. The practical implications are profound, particularly in high-stakes domains like criminal justice, lending, and hiring, where fairness is critical.

Future work could explore extending the methodology to non-linear classifiers and considering other forms of misclassification rates, such as false discovery and false omission rates. Additionally, handling smaller datasets where covariance estimates might be less reliable remains an important area for improvement.

Conclusion

The paper presents a significant advancement in fairness-aware machine learning by introducing and rigorously defining the concept of disparate mistreatment. The proposed approach offers a robust framework for building classifiers that ensure fair treatment across different social groups based on misclassification rates, paving the way for more equitable AI systems. The practical applications and theoretical underpinnings of this work will likely influence future research and development in the quest to build fairer automated decision-making systems.