FAI: A Normative Economic Approach to Fairness in AI

算法公平性方面的大量工作致力于防止人工智能（AI）加剧社会偏见。这些文献中的主要观点将公平等同于没有偏见，或者寻求在人口群体之间实现某种形式的统计平等。相比之下，该项目寻求植根于规范经济学的替代方法，该领域通过询问“应该是什么”来评估政策和项目。这项工作是由两个观察结果驱动的。首先，对个人和群体的公平可以根据人们的偏好以效用函数的形式来实现。其次，算法公平的传统概念可能与福利（群体的整体效用）不一致，包括公平标准打算保护的那些群体的福利。该项目的目标是建立规范的经济方法，作为人工智能公平性研究的核心工具。为此，该团队进行了两个研究问题。首先，规范经济学的观点能否与现有的人工智能公平方法相协调？其次，如何利用规范经济学来重新思考人工智能的公平性应该是什么？该项目将把理论和算法的进步整合到用于为难民重新安置决策提供信息的实际系统中。我们将从公平的角度审视这一制度，最终的目标是确保公平保障和福利。研究计划包括两个主要方向，基于先前的工作表明，分类器结合基于奇偶的公平性标准可能是帕累托低效的。也就是说，所有群体——包括受保护群体——的福利在一个不那么公平的分类下会更高。在第一个方向上，该项目将这种观察扩展到非凸问题，然后从处理中扩展到处理后的偏差缓解。计划中的研究还将研究多个决策者之间的相互作用及其对公平和福利等社会目标的影响。在第二个方向上，该项目发展了一种新的概念，其中分类器被视为公共资源或商品。然后，这项工作借鉴了公平分配的思想，这是规范经济学的一个长期建立的分支，定义和应用严格的公平概念，以及核心的具体概念。要将这个想法付诸实践，必须解决几个挑战：核心中存在分类器的条件、分类器的计算算法以及训练集的泛化。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

A vast body of work in algorithmic fairness is devoted to preventing artificial intelligence (AI) from exacerbating societal biases. The predominant viewpoints in this literature equates fairness with lack of bias or seeks to achieve some form of statistical parity between demographic groups. By contrast, this project pursues alternative approaches rooted in normative economics, the field that evaluates policies and programs by asking "what should be". The work is driven by two observations. First, fairness to individuals and groups can be realized according to people’s preferences represented in the form of utility functions. Second, traditional notions of algorithmic fairness may be at odds with welfare (the overall utility of groups), including the welfare of those groups the fairness criteria intend to protect. The goal of this project is to establish normative economic approaches as a central tool in the study of fairness in AI. Towards this end the team pursues two research questions. First, can the perspective of normative economics be reconciled with existing approaches to fairness in AI? Second, how can normative economics be drawn upon to rethink what fairness in AI should be? The project will integrate theoretical and algorithmic advances into real systems used to inform refugee resettlement decisions. The system will be examined from a fairness viewpoint, with the goal of ultimately ensuring fairness guarantees and welfare.The research plan includes two main directions based on previous work has shown that classifiers incorporating parity-based fairness criteria can be Pareto inefficient. That is, the welfare of all groups—including the protected group—would be higher under a classifier that is less fair. In the first direction, the project extends this observation to non-convex problems and then from in-processing to post-processing bias mitigation. The planned research will also study the interaction between multiple policy makers and its impact on social goals such as fairness and welfare. In the second direction, the project develops a new conceptualization in which classifiers are viewed as public resources or goods. This work then draws on ideas from fair division, a long-established branch of normative economics that defines and applies rigorous notions of fairness, and on the specific notion of the core. To put this idea into practice, there are several challenges that must be addressed: conditions for the existence of classifiers in the core, algorithms for their computation, and generalization from a training set.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.