报告题目：Integrate, divide, and conquer: on sparse and low-rank multivariate statistical learning

报告时间：2017年12月29日（周五） 9:00

报告地点：统计学院办公楼1层104会议室

主讲人： Dr. Kun Chen is an Assistant Professor in the Department of Statistics, University of Connecticut (UConn), and a Research Fellow at the Center for Public Health and Health Policy, UConn Health Center. Chen’s research mainly focuses on integrative multivariate learning, high-dimensional statistics, and healthcare analytics with large-scale heterogeneous data. He has extensive interdisciplinary research experience in a variety of fields, including insurance, ecology, biology, and public health. Currently he is funded by NSF for developing an integrative multivariate learning framework and for developing dimension reduction and prediction methods for heterogeneous responses; he is a co-PI in an NIH-funded data-driven suicide prevention study, which aims to improve suicide risk identification by leveraging integrated big data from disparate sources in healthcare system. Chen serves as Secretary of the newly established New England Statistical Society since 2017, and serves as an Associate Editor of Sankhya: The Indian Journal of Statistics since 2016. He has received Recognition for Teaching Excellence at UConn for multiple times since 2013.

报告摘要: Large-scale multivariate/multi-view data, or the measuring of distinct yet interrelated sets of characteristics pertaining to a single set of subjects, has become increasingly common. An integrative learning of multiple data attributes, model structures or objectives often enables us to gain extraordinary insight of complicated data generation mechanisms through utilizing information from various lenses and angles. In this talk, we present several new integrative sparse and low-rank models, crafted for conducting robust dimension reduction and outlier detection, predictive modeling with multi-view features, data fusion with incomplete and mixed-type outcomes, among others. In particular, we tackle the challenging problem of recovering a co-sparse and low-rank structure in large-scale multivariate regression. We propose divide-and-conquer strategies to isolate the recovery of each latent association pathway, for which the task simplifies greatly to a co-sparse unit-rank regression problem, permitting consistent estimation and scalable computation. The proposed methods enjoy both computational and statistical performance guarantees, and their efficacy is demonstrated by several applications in genetics, finance and public health.