统计数科学术讲座

Large-Scale Inference Via The Adaptive Projection Estimator

演讲者:Prof. Ze-Min ZHENG,University of Science and Technology of China

时间:2020-11-26 17:00-18:00

地点:Tencent Meeting ID: 919 850 940

报告简介  Abstract

Large-scale inference via de-biasing or relaxed orthogonalization has gained increasing popularity in high dimensions. Despite their usefulness, existing methods typically allow a much smaller number of nonzero coefficients compared to that in consistent estimation. To alleviate such constraint and take full advantage of the identifiable predictors, we develop a new inference procedure via the adaptive projection estimator (APE). The construction of APE is based on an adaptive orthogonalization vector which is strictly orthogonal to the identifiable covariate vectors, and at the same time being a relaxed orthogonalization against the remaining unidentifiable ones. In this way, it eliminates the bias induced by the estimation errors of identifiable coefficients and substantially enhances the inference efficiency. Furthermore, by considering approximately sparse models mixed of identifiable and weak signals, we establish comprehensive theoretical guarantees for valid inference via APE. In particular, after exploring partially penalized and penalized regressions separately, two complementary theoretical bounds are provided, showing that the maximum number of nonzero identifiable coefficients can be at least as large as that for consistent estimation. The usefulness of the proposed method is demonstrated through simulations and the stock short interest and aggregate return data set.


嘉宾简介  About the Speaker

郑泽敏,现为中国科学技术大学管理学院教授、统计与金融系主任、博士生导师,其研究方向是高维统计推断和大数据问题。郑泽敏博士在横跨这一领域的若干关键研究课题上取得了富有创造性的研究成果,研究成果发表在Journal of the Royal Statistical Society: Series B(JRSSB)、Annals of Statistics(AOS)、Operations Research(OR)、Journal of Machine Learning Research(JMLR)、Journal of Business & Economic Statistics (JBES)等国际统计学、机器学习、计量经济学及管理优化领域的顶级期刊上,曾获南加州大学授予的优秀科研奖和美国数理统计协会颁发的科研新人奖,并于2017年入选第十三批中组部‘千人计划’青年项目。他的具体研究成果既包括高维模型选择及参数渐近无偏估计等普遍适用的统计学习方法论,也包含针对大数据问题而提出的具有计算高效性的新型统计方法,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管理学热点问题中具有重要的应用价值。


讲座海报 Poster

2. 统计数科学术讲座-郑泽敏教授.jpg