设为首页 | 加关注 | 设为收藏

通知公告

Notice

China's income distribution research center

首页 通知公告

通知公告

林伟教授学术讲座公告
来源:统计与数学学院    编辑:佚名    时间:2016/11/24    点击数:
林伟教授学术讲座公告 报 告 人:林 伟(北京大学) 报告时间:11月30号下午2:30—3:30 (报告一) 12月一号上午10:00—11:00(报告二) 报告地点:文波楼4楼会议室 题目一: Randomizing the Simplex: Scalable and Accurate Inference for Clustered Categorical Data 摘要一:In this talk, we are concerned with a type of complex data structure that arises in diverse areas of application such as metagenomics, text analysis, bibliometrics, and marketing science. We first review two schools of modeling and analysis techniques for such data. In particular, when the data are viewed as clustered categorical data, we propose a high-dimensional logistic-normal multinomial model and develop methods for parameter estimation. To mitigate the computational intractability, a stochastic approximation EM algorithm with Hamiltonian Monte Carlo sampling is introduced, which is highly scalable and accurate. We further consider condition-number-type regularization on the covariance structure of random effects. Algorithmic convergence and risk properties are established, and the advantages of our method are demonstrated on simulated and real data. 题目二: Causal Inference with Many Invalid Instruments 摘要二:Instrumental variables methods have been widely used to estimate the causal effects of exposure variables on an outcome of interest. In genetical genomics, genetic variants are often used as high-dimensional instruments to control for unmeasured confounding. Existing methods for dealing with high-dimensional instruments, however, either require the unrealistic assumption on the validity of all instruments or can only handle a single exposure variable. Motivated by such emerging applications, we investigate the identifiability of high-dimensional instrumental variables models in the presence of possibly invalid instruments, and propose a two-stage regularization framework for estimating the causal effects. We further develop methods for obtaining debiased estimates that are asymptotically normal, which allow us to construct valid confidence intervals for the causal effects. The proposed procedure is efficiently implemented, and their theoretical properties are established and verified through simulation studies. The usefulness of our approach is demonstrated on a mouse obesity dataset. 报告人简介: 林伟教授,2011年8月于University of Southern California获得博士学位。2011年9月-2014年6月在University of Pennsylvania从事博士后研究。 2014年加入北京大学,并入选国家‘青年千人’计划。林伟教授研究兴趣涉及高维数据分析、大数据、因果推断、生存分析及其在基因和生物数据等领域的应用。主持国家重点研究计划项目,国家自然科学基金面上项目等多项基金,在Journal of American Statistical Association, Biometrika等统计学国际顶级学术期刊发表论文10余篇。长期担任国际顶级期刊 Annals of Statistics,Journal of American Statistical Association,Biometrika,Journal of Machine Learning Research,IEEE Transactions on Signal Processing审稿人。
copyright©2012 znufe.edu.cn.all rights reserved. 地址:湖北省武汉市东湖高新技术开发区南湖大道182号 邮编:430073 电话:027-88986475 鄂ICP备05003344号