CAREER: Towards Long-term Fairness in Sequential Decision Making

公平机器学习是一个旨在减少机器自动化决策中的歧视和偏见的研究课题，是使包括自动决策系统在内的人工智能系统大规模部署得到广泛社会接受的关键之一。目前，大多数关于公平机器学习的研究都是基于静态设置，机器模型在部署后只对每个个体做出一次决策。然而，在实际情况下，机器学习模型通常会被部署在一段时间内做出连续的决策。在这种顺序决策设置中，确保每一步的公平性并不能保证长期的公平性，这给公平机器学习社区提出了一个具有挑战性和紧迫性的问题，即如何实现长期公平性。该项目将通过极大地推进对动态环境中公平的基本问题的理解，为解决不一致的公平概念之间的冲突提供途径，并为长期公平机器学习的有限知识基础做出贡献，这对于许多现实世界的应用是必不可少的。该教育项目将涉及本科生、毕业生和高中生，以提高他们在机器学习和人工智能领域解决问题的知识和技能，并吸引学生，特别是那些来自代表性不足群体的学生，从事STEM领域的职业。该项目将利用Pearl的结构因果模型为长期公平的机器学习奠定基础。研究者将专注于连续的决策设置，过去做出的决定可能会对未来的数据产生影响。软干预将被用来捕捉决策模型部署的因果效应。研究者将开发基于因果模型的长期公平性的通用公式，以便它可以通过因果推理技术来测量。然后，根据决策者是否有足够的历史数据，研究者将通过三个渐进式研究任务来研究离线和在线学习设置：(1)研究在给定足够的历史训练数据的情况下实现长期公平的策略和算法；(2)研究如何在捕捉历史动态的同时，对未来数据进行预测，使所建立的决策模型在可预测的未来具有公平性；(3)转向在线学习，决策者只有很少或没有训练数据，但可以在线更新决策模型，并希望最终实现公平。最后，研究了两种现实的扩展，包括不可识别情况和半马尔可夫模型。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。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2)Fair machine learning, a research topic that aims to reduce discrimination and bias in machine-automated decisions, is one of the keys for enabling broad societal acceptance of large-scale deployments of AI systems including automated decision-making systems. Currently the majority of studies in fair machine learning are based on static settings where the machine model makes the decision only once for each individual after its deployment. However, in practical situations, the machine learning model will usually be deployed to make sequential decisions over a period of time. In such sequential decision-making settings, ensuring fairness for each single step does not guarantee fairness in the long-term, presenting a challenging and urgent problem to the fair machine learning community about achieving long-term fairness. This project will make a transformative change to fair machine learning by greatly advancing the understanding of fundamental issues of fairness in dynamic settings, shedding light on the path to addressing conflicts between inconsistent fairness concepts, and contributing to the limited base of knowledge in long-term fair machine learning which is imperative for many real-world applications. The education program will involve undergraduates, graduates and high school students to enhance their knowledge and skills in solving problems in machine learning and artificial intelligence, and attract students especially those from underrepresented groups to pursue careers in STEM.This project will set up the foundation for long-term fair machine learning by leveraging Pearl's Structural Causal Model. The investigator will focus on the sequential decision-making setting where decisions made in the past may have an impact on future data. Soft intervention will be utilized to capture the causal effect of the deployment of decision models. The investigator will develop universal formulations for the long-term fairness based on the causal model so that it can be measured by causal inference techniques. Then, depending on whether the decision maker has access to adequate historical data, the investigator will study both offline and online learning settings via three progressive research tasks: (1) to study strategies and algorithms for achieving long-term fairness given sufficient historical training data; (2) to study how to not only capture the dynamics in the history but also predict the data in the future so that the decision model built would be fair in the predictable future; and (3) to move on to online learning where the decision maker has few or no training data but could update the decision model in an online manner and wants to achieve fairness eventually. Finally, two realistic extensions including unidentifiable situations and semi-Markovian models will be studied.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.