TRIPODS: Data Science for Improved Decision-Making: Learning in the Context of Uncertainty, Causality, Privacy, and Network Structures

研究人员提议建立一个数据科学中心，将计算机科学、信息科学、数学、运筹学和统计学的专业知识结合起来，以改进决策。他们的目标是从事有助于数据科学理论基础的基础研究。所选择的研究主题具有造福整个社会的应用，并整合了该项目汇集的各学科的观点。提出了五个具体的研究方向：隐私与公平、社交图学习、学会干预、不确定性量化和深度学习。该中心的目的是增进这些领域的知识，拓宽能够对这些具有挑战性的问题作出贡献的学科和视角的范围。研究人员计划通过在线研讨会、研讨会和学生会议将康奈尔以外的社区纳入其中。这些研究成果将为数据科学在几个对社会具有重要意义的主题领域提供迫切需要的基础。由于该中心位于多个学科的交叉点，其智力优势横跨所有涉及的学科，研究结果可能会转化为每一个学科的新算法和方法。研究重点涵盖五个核心领域。1.隐私和公平。随着数据科学在社会的许多领域变得无处不在，随着它越来越多地被用于帮助敏感领域的决策，通过保障隐私和公平来保护个人变得至关重要。调查人员建议研究提供这种保证的理论基础，并揭示其固有的局限性。2.学习社交图。在将数据科学应用于个人和更大的社会系统之间的互动时，许多基本问题都涉及支撑个人之间联系的社交网络。研究人员将开发新的技术来理解这些网络的结构和在其中发生的过程。学会干预。学习好的干预措施(包括政策、建议和治疗)的数据驱动的方法激发了关于序贯实验设计、反事实推理和因果推理的基础的挑战性问题。不确定度量化。量化特定预测或结论的不确定性是数据科学的一个关键需求，特别是当应用于对人类受试者具有潜在后果的决策时。研究人员将开发统计工具和理论保证，以评估数据科学中流行算法做出的预测的不确定性。5.深度学习。深度学习算法在实际应用中取得了令人印象深刻的进展。尽管人们很好地理解了他们的基本构件，但对于他们学到了什么以及为什么他们能如此好地概括，仍然存在模棱两可的问题。有迹象表明，他们可能会学习数据流形，并且优化算法的类型会影响泛化。在我们对这些现象的理论理解上的进步需要优化、统计学和数学的共同努力，但可能导致对数据科学的所有方面的洞察。该项目的资金来自CEISE计算和通信基金会、MPS数学科学部、MPS多学科活动办公室和增长融合研究。(融合可以被描述为深度整合来自多个领域的知识、技术和专业知识，以形成应对科学和社会挑战和机遇的新的和扩大的框架。该项目通过将代表包括数学、统计学和理论计算机科学在内的许多学科的社区聚集在一起，并让将数据科学应用于实际研究问题的社区参与进来，来促进融合。)

The researchers propose to create a center of data science for improved decision-making that combines expertise from computer science, information science, mathematics, operations research, and statistics. Their goal is to pursue basic research that will contribute to the theoretical foundations of data science. The research topics chosen have applications that can benefit society as a whole and integrate the perspectives of the disciplines that the project brings together. The five concrete research directions proposed are: Privacy and Fairness, Learning on Social Graphs, Learning to Intervene, Uncertainty Quantification, and Deep Learning. The aim of the Center is to advance knowledge in these areas and to broaden the range of disciplines and perspectives that can provide contributions to these challenging issues. The researchers plan to incorporate the community beyond Cornell through online seminars, workshops, and student conferences. The research findings will provide an urgently needed foundation for data science in several topic areas of importance to society. As the center is placed at the intersection of multiple disciplines, the intellectual merit spans all disciplines involved and findings may translate to new algorithms and approaches in each one of them. The research focus spans five core areas. 1. Privacy and Fairness. As data science becomes pervasive across many areas of society, and as it is increasingly used to aid decision-making in sensitive domains, it becomes crucial to protect individuals by guaranteeing privacy and fairness. The investigators propose to research the theoretical foundations to providing such guarantees and to surface inherent limitations. 2. Learning on Social Graphs. Many of the fundamental questions in applying data science to the interactions between individuals and larger social systems involve the social networks that underpin the connections between individuals. The researchers will develop new techniques for understanding both the structure of these networks and the processes that take place within them.3. Learning to Intervene. Data-driven approaches to learning good interventions (including policies, recommendations, and treatments) inspire challenging questions about the foundations of sequential experimental design, counterfactual reasoning, and causal inference.4. Uncertainty Quantification. Quantifying uncertainty about specific predictions or conclusions represents a key need in data science, especially when applied to decision-making with potential consequences to human subjects. The researchers will develop statistical tools and theoretical guarantees to assess the uncertainties of predictions made by popular algorithms in data science. 5. Deep Learning. Deep Learning algorithms have made impressive advances in practical settings. Although their basic building blocks are well understood, there is still ambiguity about what they learn and why they generalize so well. There are indications that they may learn data manifolds and that the type of optimization algorithm influences generalization. Advances in our theoretical understanding of these phenomena requires combined efforts from optimization, statistics, and mathematics but could lead to insights for all aspects of data science.Funds for the project come from CISE Computing and Communications Foundations, MPS Division of Mathematical Sciences, MPS Office of Multidisciplinary Activities, and Growing Convergent Research. (Convergence can be characterized as the deep integration of knowledge, techniques, and expertise from multiple fields to form new and expanded frameworks for addressing scientific and societal challenges and opportunities. This project promotes Convergence by bringing together communities representing many disciplines including mathematics, statistics, and theoretical computer science as well as engaging communities that apply data science to practical research problems.)