高维复杂数据分析中具有可重复性的统计学习方法研究及其应用

高维复杂数据分析是当前大数据与人工智能国家战略的重要组成部分，同时也是国际统计学及机器学习研究的热点领域，在动态定价策略、产品组合优化、平台推荐系统和社交网络推断等当代管理学热点问题中有着广泛应用。然而，由于样本中的随机噪音及高维数据的复杂结构，数据分析的结果在不同的样本下常常不具有可重复性，这给统计学习中结论的可靠性带来极大挑战。本项目拟基于在模型中引入人工控制变量构造检验统计量的思想，发展若干统计学习新方法以控制错误发现率，从而实现复杂数据分析中科学发现的可重复性。主要研究工作将从多响应回归问题中的错误发现率控制、带删失响应数据中统计学习的可重复性、多变点检测问题中具有可重复性的统计推断这三个方向展开。相关研究成果将应用于药品智慧评价与管理等实际问题。

High-dimensional complex data analysis is an intrinsic component of the current strategy “Big Data and Artificial Intelligence” in our country, also the forefront of statistical science and machine learning, with wide applications in hot issues of modern management such as dynamic pricing, assortment optimization, recommender systems, and social network inference. However, due to the random noises in the sample and the complex structures of high-dimensional data, the findings of data analysis may not be reproducible, which poses great challenges to the reliability of statistical learnings. In this project, we will utilize the idea of false discovery rate control through control variables to innovate several new statistical learnning methods, so that the scientific findings in complex data analysis can be reproducible. The main research contents include three aspects: false discovery rate control in multiple response regressions, reproducible statistical learning with censored responses，and reproducible change point detection in multiple change point problems. The new methods aim at real applications including intelligent evaluation and management of drugs.