CAREER: A Stable Foundation for Trustworthy Data Analysis

每天都有大量的数据被收集、分析并用于做出高风险的决策，这就提出了许多关于如何以值得信赖的方式使用这些数据的问题。 这个项目是关于两个这样的问题：（1）研究人员如何防止错误的发现，并使用数据来了解人口的有意义的事实，而不过度拟合该数据？ 尽管几十年来一直在研究防止错误发现的方法，但它仍然是科学界的一个令人烦恼的问题。 (2)研究人员如何使用有价值但敏感的数据来了解一个群体，而不损害该数据中个人的隐私？ 这项任务已经被证明是相当微妙的，并且已经有几次针对所谓的匿名数据集的高调攻击，导致对最常用的方法缺乏信心。 虽然它们看起来似乎无关，但令人惊讶的是，这两个问题都可以使用稳定的算法来解决--对输入中的微小变化不敏感的算法。 在过去的十年中，差分隐私作为一种强大的算法稳定性形式出现，它保证了高度的个人隐私，但也允许高度准确的数据分析。 最近，差异隐私已经被证明可以防止交互式数据分析中的错误发现--重复分析同一数据集的常见情况，这与“科学中的统计危机”有关。“该项目将采取统一的方法，通过算法稳定性来推进隐私和虚假发现方面的最新技术。 该项目的主要成果将是建立交互式数据分析的理论基础，为这些领域的核心问题开发新的计算效率高的稳定算法，在理论和实践中理解隐私和交互式数据分析的局限性，并扩大算法稳定性的范围，以应对可信数据分析中的其他挑战。

Every day, massive amounts of data are collected, analyzed, and used to make high-stakes decisions, raising many questions about how to use this data in a trustworthy manner. This project is about two such questions: (1) How can researchers prevent false discovery, and use data to learn meaningful facts about a population without overfitting to that data? Despite decades of research into methods for preventing false discovery, it remains a vexing problem for the scientific community. (2) How can researchers use valuable but sensitive data to learn about a population without compromising the privacy of individuals in that data? This task has proven to be quite delicate, and there have been several high profile attacks on supposedly anonymous datasets, causing a lack of confidence in the most commonly used approaches. Although they may seem unrelated, surprisingly, both of these questions can be addressed using stable algorithms---algorithms that are insensitive to small changes in their inputs. In the past decade, differential privacy emerged as a strong form of algorithmic stability that guarantees a high degree of individual privacy, yet admits highly accurate data analysis. More recently, differential privacy has been shown to prevent false discovery in interactive data analysis---the common scenario where the same dataset is analyzed repeatedly, which has been implicated in a "statistical crisis in science."This project will take a unified approach to advancing the state-of-the-art in privacy and false discovery via algorithmic stability. The main outcomes of this project will be building the theoretical foundations of interactive data analysis, developing new computationally efficient stable algorithms for central problems in these areas, understanding the limits of privacy and interactive data analysis both in theory and in practice, and broadening the reach of algorithmic stability to address other challenges in trustworthy data analysis.