Collaborative Research: AF: Medium: Machine Learning Markets: Dynamics, Competition, and Interventions

协作研究：AF：媒介：机器学习市场：动态、竞争和干预

• 批准号: 2312775
• 负责人: Lillian Ratliff
• 金额: $ 44.99万
• 依托单位: University of Washington
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2312775&HistoricalAwards=false
• 关键词: Collaborative; Research; AF; Medium; Machine

Nearly every aspect of modern life involves predictions of machine learning (ML) models: e.g., where we choose to shop, live, or apply for jobs. ML models are built and maintained by service providers, who use predictions to offer services to individuals. Based on prediction quality, individuals choose amongst services, or may choose to use none. This interaction gives rise to an ML-enabled market wherein the decisions made by providers and individuals are highly consequential. As individuals choose amongst services, such services gain or lose not only customers, but also access to data that allows them to refine their predictions. The dynamics of interactions between providers and users with varied motivations are more complex than traditional market models. This project aims to develop the theoretical and algorithmic foundations for characterizing and shaping ML-enabled market conditions and algorithmic tools for achieving improved social outcomes. Beyond the theoretical contributions, products of this project will support the development of policy for governing ML-enabled markets in a fair and equitable manner. The project's educational goals include mentoring researchers at the undergraduate and graduate levels, developing new course materials, and promoting dialog on equity in ML through talks, reading groups, and topical workshop activities coordinated with the NSF Institute for Foundations of Data Science at University of Washington.Interactions between individuals, represented by data, and providers, represented by prediction functions in ML-enabled markets, give rise to complex dynamics, and competitive or cooperative games. The choices of both providers and individuals may be strategic or myopic, depending on whether agents anticipate how their choices will affect future market conditions. Providers and users may act according to a variety of objectives: predictive accuracy (service quality), market share, privacy, fairness, or even adversarial intent. This project analyzes the complex interactions between providers and individuals in ML-enabled markets, with a research agenda comprised of three thrusts: (1) Characterize the participation game and dynamics that arise when users with a variety of objectives act strategically, while providers use data myopically to improve the accuracy of their predictions; (2) Characterize the prediction--retention game and dynamics that arise between strategic providers (including the incentives and social costs of strategic behavior), when users choose their participation level myopically; (3) Combine insights from the first two threads to design algorithmic interventions that improve metrics of outcomes such as social welfare and fairness in ML-enabled markets. Carrying out this agenda will entail developing new theory and algorithms at the intersection of game theory, statistical learning, dynamical systems, and optimization. Challenges due to nonlinear dynamics, nonconvex landscapes, and information limitations will be addressed, and the equilibrium landscape of competitive games with novel structure will be characterized, contributing to the core fields of game theory and machine learning, and their social impact.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.

现代生活的几乎每个方面都涉及到机器学习（ML）模型的预测：例如，我们选择在哪里购物、居住或申请工作。机器学习模型由服务提供者构建和维护，他们使用预测向个人提供服务。基于预测质量，个人可以在服务中进行选择，也可以选择不使用。这种交互产生了一个支持机器学习的市场，在这个市场中，供应商和个人做出的决定是非常重要的。当个人在服务中进行选择时，这些服务不仅会获得或失去客户，还会获得允许他们改进预测的数据。具有不同动机的供应商和用户之间的互动动态比传统的市场模型更为复杂。该项目旨在发展理论和算法基础，以描述和塑造机器学习支持的市场条件和算法工具，以实现改善的社会结果。除了理论贡献之外，该项目的产品将支持以公平和公平的方式管理机器学习市场的政策发展。该项目的教育目标包括指导本科生和研究生水平的研究人员，开发新的课程材料，并通过与华盛顿大学NSF数据科学基础研究所协调的讲座、阅读小组和专题研讨会活动，促进机器学习公平性的对话。在机器学习支持的市场中，以数据为代表的个体和以预测功能为代表的提供者之间的互动产生了复杂的动态，以及竞争或合作的游戏。供应商和个人的选择可能是战略性的，也可能是短视的，这取决于代理商是否预测到他们的选择将如何影响未来的市场状况。提供者和用户可以根据各种目标采取行动：预测准确性（服务质量）、市场份额、隐私、公平，甚至是敌对意图。该项目分析了机器学习市场中供应商和个人之间的复杂互动，其研究议程包括三个重点：(1)描述了当具有各种目标的用户采取战略行动时出现的参与游戏和动态，而供应商则短视地使用数据来提高其预测的准确性；(2)描述当用户短视地选择他们的参与水平时，战略提供者之间出现的预测-留存博弈和动态（包括战略行为的激励和社会成本）；(3)结合前两条线索的见解，设计算法干预措施，改善基于机器学习的市场中的社会福利和公平等结果指标。执行这一议程将需要在博弈论、统计学习、动力系统和优化的交叉点上发展新的理论和算法。将解决非线性动力学、非凸景观和信息限制带来的挑战，并描述具有新结构的竞争博弈的均衡景观，为博弈论和机器学习的核心领域及其社会影响做出贡献。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。