报告题目：Two Types of Parameters in Subsampling for Massive Data

报告时间：2026年7月17日下午14:00

报告地点：南湖校区新图书馆5302室

主办单位：红杏视频

报告人：王海鹰

报告人简介：Haiying Wang is an Associate Professor of Statistics at the University of Connecticut, with a Ph.D. from the University of Missouri (2013) and an M.S. from the Chinese Academy of Sciences (2006). His research covers informative subdata selection for big data, model selection and averaging, measurement error models, optimal design, and semi-parametric regression. His work has been published in leading journals and presented at premier conferences, including Biometrika, IEEE Transactions, Journal of the American Statistical Association, Journal of Machine Learning Research, ICML, and NeurIPS.

摘要：As massive datasets become the norm, subsampling has emerged as a crucial technique to make statistical computation feasible. This talk introduces the data-dependent subsampling approach, focusing on the critical distinction between two types of targets in subsampling: approximating the full-data estimator versus estimating the true population parameter. When the goal is to approximate a computationally intractable full-data estimator, conditional distributions are sufficient, and Inverse Probability Weighting (IPW) combined with optimal design (e.g., OSMAC) provides a robust solution, even under model misspecification. Conversely, when the goal is to infer the true population parameter, the unconditional distribution of the data becomes relevant, and unweighted approaches—such as likelihood-based methods—often yield significantly higher statistical efficiency, provided the model is correctly specified. We will discuss the theoretical properties of both approaches and explore their connections and differences. We will also point out some interesting facts regarding subsampling.