[논문 리뷰] The Frontiers of Fairness in Machine Learning

Alexandra Chouldechova, Aaron Roth|arXiv (Cornell University)|2018. 10. 20.

Ethics and Social Impacts of AI참고 문헌 13인용 수 242

한 줄 요약

2018 CCC visioning workshop의 공정성에 대한 조사로 머신 러닝에서의 공정성을 조사하고, 현재의 정의, 도전 과제 및 알고리즘적 공정성의 기초 이해를 위한 유망한 연구 방향을 개략한다.

ABSTRACT

The last few years have seen an explosion of academic and popular interest in algorithmic fairness. Despite this interest and the volume and velocity of work that has been produced recently, the fundamental science of fairness in machine learning is still in a nascent state. In March 2018, we convened a group of experts as part of a CCC visioning workshop to assess the state of the field, and distill the most promising research directions going forward. This report summarizes the findings of that workshop. Along the way, it surveys recent theoretical work in the field and points towards promising directions for research.

연구 동기 및 목표

• Identify and articulate the fundamental questions and definitions shaping the science of fairness in machine learning.
• Survey major theoretical approaches to statistical and individual notions of fairness and their limitations.
• Highlight open problems and promising directions to build a robust, compositional, and dynamic theory of fairness.

제안 방법

• Review and synthesize theoretical work on fairness from statistical and individual perspectives.
• Discuss causes of unfairness such as data bias, distributional differences, and exploratory data needs.
• Examine fairness definitions, impossibility results, and computational considerations.
• Explore advances in fair representations and their limitations under potential adversarial recovery of sensitive information.
• Discuss fairness beyond static, one-shot classification, including dynamics, composition, and sequential decision making.

실험 결과

연구 질문

• RQ1 What are the main statistical and individual notions of fairness and their tradeoffs?
• RQ2 Can we obtain practical fairness guarantees that extend beyond a few protected groups to many or infinite groups?
• RQ3 How do fairness notions behave under composition of multiple components or in dynamic environments?
• RQ4 How can data bias be modeled, corrected, or mitigated without sacrificing overall performance?
• RQ5 What roles do fair representations and non-classification settings (bandits, reinforcement learning) play in achieving fairness?

주요 결과

• Statistical fairness notions (parity, equalized odds, predictive value) can conflict, and simultaneous satisfaction is often impossible.
• Fairness definitions can be computationally hard to enforce, yet practical algorithms exist for certain settings.
• Fair representation learning and adversarial approaches can mitigate disparate impact but may be vulnerable to information recovery by stronger downstream models.
• Bias and feedback loops in data generation and collection can propagate or amplify unfairness, necessitating careful data correction and modeling of construct spaces.
• Fairness across individuals and across repeatedly interacting system components (composition) remains challenging, with limited understanding of long-term dynamics and incentives.