On Formalizing Fairness in Prediction with Machine Learning (1710.03184v3)

Abstract: Machine learning algorithms for prediction are increasingly being used in critical decisions affecting human lives. Various fairness formalizations, with no firm consensus yet, are employed to prevent such algorithms from systematically discriminating against people based on certain attributes protected by law. The aim of this article is to survey how fairness is formalized in the machine learning literature for the task of prediction and present these formalizations with their corresponding notions of distributive justice from the social sciences literature. We provide theoretical as well as empirical critiques of these notions from the social sciences literature and explain how these critiques limit the suitability of the corresponding fairness formalizations to certain domains. We also suggest two notions of distributive justice which address some of these critiques and discuss avenues for prospective fairness formalizations.

• The paper surveys a range of fairness formalizations, evaluating their ability to prevent algorithmic bias in high-impact sectors.
• It contrasts methods like counterfactual and group fairness, detailing their trade-offs and practical limitations.
• The study introduces new theories such as Equality of Resources to inspire holistic interdisciplinary approaches for fairer ML predictions.

Formalizing Fairness in Machine Learning Prediction Algorithms

The paper "On Formalizing Fairness in Prediction with Machine Learning" by Pratik Gajane and Mykola Pechenizkiy addresses the critical topic of fairness in machine learning algorithms used in high-impact prediction applications. The authors survey various approaches to formalizing fairness in these algorithms and discuss the conceptual underpinnings from social sciences literature to present a comprehensive critique of existing fairness formalizations. This paper meticulously explores the intersection between machine learning and social theories of distributive justice to provide insights into the strengths and limitations of current fairness paradigms.

The discussion commences with an examination of the implications of discrimination in machine learning systems, particularly in sectors like credit, employment, education, and criminal justice. The significant challenge is to ensure that machine learning algorithms do not perpetuate or amplify societal biases against protected demographic groups. The paper rightly identifies that no single formalization of fairness has emerged as a consensus standard, hence the necessity of examining different formalizations through the lenses of both theoretical and empirical critiques.

Existing Formalizations of Fairness

The authors categorize fairness measures into several distinct formalizations such as:

1. Fairness through Unawareness: This formalization assumes fairness can be achieved by not explicitly using protected attributes in the prediction process. However, it is critiqued for its inadequacy in addressing discrimination when protected attributes can be deduced from other available data.
2. Counterfactual Measures: These employ causal inference to ensure fairness by comparing the actual and counterfactual outcomes of predictions. While innovative, they are scrutinized for potential biases like hindsight bias, which question their reliability in domains with pre-existing systemic biases.
3. Group Fairness: Also known as statistical parity, this measure demands equal prediction outcomes across groups. This approach is particularly relevant when reliable label information is not available, although it has been critiqued for possibly disadvantaging more qualified individuals within groups.
4. Individual Fairness: This paradigm posits that similar individuals should receive similar predictions. The main deficiency lies in its dependency on a suitable distance metric, which, if biased, can lead to unfair outcomes.
5. Equality of Opportunity: This formalization requires equal true positive rates across groups. It relates closely to societal notions of distributive justice but may overlook the broader societal factors that shape individuals' opportunities.
6. Preference-based Fairness: This involves maximizing the group benefit through preference satisfaction. However, complexities arise in achieving envy-free or Pareto-efficient solutions, which makes it challenging to apply universally.

Prospective Fairness Formalizations

In addition to reviewing existing strategies, the authors propose two prospective theories — Equality of Resources and Equality of Capability of Functioning. Both theories seek to address the social and natural endowment disparities by considering broader societal factors. Although these theories present their own conceptual difficulties, particularly in measurable implementation, they offer an intriguing direction for future research that aligns with more holistic societal considerations.

Conclusion and Future Directions

The authors conclude by emphasizing the critical need for interdisciplinary approaches bringing insights from social sciences into machine learning fairness formalizations. By highlighting the inadequacies in existing paradigms and advocating for new avenues, this work sets the stage for further refinement in the field. Future directions might include developing robust methodologies to integrate social considerations and untangle complex societal biases within algorithmic systems, allowing for the continued evolution of fairer and more just machine learning prediction algorithms.

Overall, the paper serves as both a survey and a clarion call to the research community to evolve fairness measures that are both technically sound and socially aware, fostering equity in machine learning applications.