AF: Small: Foundations for Societal Machine Learning

机器学习已经成为一种非常成功和实用的工具，用于理解数据、启用新技术和帮助人类决策。然而，它在影响人们的应用程序中的使用越来越多，也引起了许多关注。这些问题包括对决策公平性的关注，对产生的激励和战略行为对这些系统准确性的影响的关注，以及对分类决策对社会福利的影响的关注。该项目旨在开发理论框架，为解决这些问题的机器学习系统奠定基础。特别是，这项工作的高级目标是能够为使用这些系统的人和受其决策影响的人提供干净的保证。具体而言，本项目围绕三个主要研究方向展开。第一个是推进对机器学习和算法背景下公平性的理解，重点是公平性条件和有偏见的训练数据之间的相互作用，以及在多阶段决策系统中实现公平性条件。第二个方向涉及战略分类，这是对具有在有限范围内修改其可观察特征的能力的代理进行分类决策的问题，如果这会导致他们喜欢的决策，那么谁可能会这样做。这项工作将解决在这种情况下具有可证明的准确性保证的算法设计中的一些基本问题，特别是对于在线顺序决策的挑战性情况。第三个方向涉及对社会福利的影响，以及将经典的准确性目标与涉及激励社会有益行为的目标相结合的学习算法的发展。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。

Machine learning has become a highly successful and practical tool for understanding data, enabling new technologies, and aiding human decision-making. However, its increased use in applications that impact people has also led to a number of concerns. These include concerns about the fairness of decisions made, concerns about incentives generated and the effect of strategic behavior on accuracy of these systems, and concerns about the impact of classification decisions on societal welfare. This project aims to develop theoretical frameworks that advance the foundations for machine learning systems that address these concerns. In particular, the high-level goal of this work is to be able to provide clean guarantees both to those using these systems and to those affected by the decisions they make.Specifically, this project is centered around three main research directions. The first is to advance the understanding of fairness in machine-learning and algorithmic contexts, with emphasis on the interaction between fairness conditions and biased training data, and on implementing fairness conditions in multi-stage decision systems. The second direction involves strategic classification, which is the problem of making classification decisions on agents that have the ability to modify their observable features to a limited extent, and who may do so if it leads to a decision they prefer. This work will tackle a number of fundamental problems in the design of algorithms with provable accuracy guarantees in such settings, especially for the challenging case of online sequential decision-making. The third direction involves impacts on societal welfare, and development of learning algorithms that combine classic accuracy goals with goals that involve incentivizing societally-beneficial behaviors.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.