CRII: III: Rethinking Fairness: Fairness as a Survival Analysis

在机器学习界内外，人们越来越担心人工智能(AI)面临偏见和歧视危机，迫切需要AI系统纳入公平约束。美国国会已经认识到这一问题，并一直在努力通过算法问责法案。它要求评估系统的准确性、公平性、偏见、歧视、隐私和自动化系统内的安全性，并要求公司纠正在此过程中发现的任何问题。大多数现有的评估公平性的工作都假设记录的可用性，其中源数据用应用公平性定义和公平性算法所需的类别进行注释。然而，这一假设在现实世界、社会敏感的各种应用中是不切实际的，从精密医学到营销分析、精算分析和累犯率预测工具。因此，迫切需要研究在实验室中设计“公平”模型与在现实世界中部署之间的差距所产生的问题。为此，这个项目将重新审视公平的基本定义，并揭示现有公平文献中源于假设实际中无法获得的信息的特质。下一步，该项目将致力于弥合目前的人工智能公平研究与现实世界部署之间的差距，从而更好地理解人工智能的社会影响，并显著减少其潜在的社会歧视。为了实现这一目标，该项目将制定一个新的公平作为生存分析问题，其中类标签的可用性并不总是得到保证，但仍然要求类似的个人得到类似的对待。第一个研究目标是从两个不同的角度来量化缺失标签情况下的个体不公平。具体地说，第一个轨道将公平性视为输入和输出空间中的相似性的相关性，这使得能够定义可用于统计审查数据的公平性度量。第二个定义将构成从稳健性角度产生的另一个公平问题，即评估类似的个人是否遭受不同水平的预测稳定性。第二个研究目标将进行初步调查，共同解决模型构建中的偏差减少和统计审查管理问题，以确保效用最大化，同时最小化个体之间的偏差。这些标准将被制定为联合优化的正规化条件，不要求所有个人都有类别标签。预计该项目的结果将包括确保在各种现实世界中对社会敏感的应用程序的公平性保证的多功能人工制品。此外，该项目将引入一个新的任务设置，为未来在人工智能公平的实际应用方面的研究铺平道路。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

There has been increasing concern within the machine learning community and beyond that Artificial Intelligence (AI) faces a bias and discrimination crisis, urgently requiring AI systems to incorporate fairness constraints. The US Congress has recognized this issue and has been trying to pass the Algorithmic Accountability Act. It demands systems be evaluated for “accuracy, fairness, bias, discrimination, privacy and security within automated systems and companies would be required to correct any issues they uncovered during the process.” Most existing work on evaluating fairness assumes the availability of records in which the source data is annotated with categories needed to apply the fairness definition and fairness algorithm at hand. This assumption, however, is impractical in a diversity of real-world, socially-sensitive applications, ranging from precision medicine to marketing analytics, actuarial analysis and recidivism prediction instruments. There is thus a critical need to study the problem that arises from the gap between the design of a “fair” model in the lab and its deployment in the real world. To this end, this project will revisit the foundational definitions of fairness and reveal idiosyncrasies in the existing fairness literature stemming from assuming information that is not available in practice. Next, this project will aim to bridge the gap between current AI fairness studies and their real-world deployment, leading to improved understanding of the societal impact of AI and significant reduction in its potential for social discrimination. To achieve this goal, the project will formulate a new fairness-as-a-survival-analysis problem, where the availability of class labels is not always guaranteed, but there is still a requirement that similar individuals are treated similarly. The first research objective focuses on quantifying individual unfairness in the presence of missing labels from two different perspectives. Specifically, the first track will see fairness as the correlation of similarity in the input and output spaces, which enables defining a fairness measure usable on statistically censored data. The second definition will constitute another fairness issue arising from the perspective of robustness, evaluating whether similar individuals suffer dissimilar levels of prediction stability. The second research objective will make an initial investigation jointly addressing bias reduction and statistical censoring management in model building, so as to ensure utility maximization while minimizing bias across individuals. These criteria will be formulated as regularization terms for joint optimization and will not require all individuals to have a class label. The outcomes of this project are expected to include versatile artifacts that ensure fairness guarantees in various real-world socially-sensitive applications. Furthermore, the project will introduce a new task setting, paving the way for future research in the practical application of AI fairness.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.