报告题目：Statistical Learning for Heterogeneous Multivariate Data

主讲人：宋心远教授（香港中文大学）

时间：2026年10月17日（周六）15:00 p.m.

地点：北院卓远楼305会议室

主办单位：统计与数学学院

摘要：

In this talk, we introduce a unified framework for modeling multivariate heterogeneous data by integrating latent variable models, mixed membership structures, and longitudinal trajectory analysis. We first introduce a mixed membership latent variable model that combines exploratory factor analysis with flexible subpopulation modeling, enabling simultaneous learning of latent factors, factor structures, and profile memberships while capturing individual heterogeneity and addressing multicollinearity. We then present a multi-label growth curve change-point model for multivariate longitudinal data, which accommodates between-individual, between-variable, and within-subject temporal heterogeneity through domain-specific latent classes and subject-specific change points, allowing individuals to exhibit distinct progression patterns across multiple outcome domains. Together, these approaches provide a general paradigm for statistical learning in heterogeneous multivariate data, emphasizing interpretable latent representations, partial memberships, and dynamic structures. Theoretically, we show that the proposed methods accurately recover the underlying heterogeneity structure. Applications to Parkinson’s disease data demonstrate their ability to uncover meaningful latent traits, domain-specific progression patterns, and refined disease subtypes, offering valuable insights for complex data analysis and precision medicine.

主讲人简介：

宋心远是香港中文大学统计与数据科学系的教授。她目前是中国教育部“长江学者”讲席教授、数理统计学会（IMS）会士（Fellow）以及国际统计学会（ISI）当选会员。她的研究兴趣涵盖潜变量模型、贝叶斯方法、生存分析、非参数与半参数方法以及统计计算,已在国际知名期刊发表论文两百余篇并担任或曾担任多个统计学与心理计量学领域国际期刊的副主编，包括：《美国统计学会会刊》（JASA）、《生物计量学》（Biometrics）、《医学统计学》（Statistics in Medicine）、《英国皇家统计学会会刊》（Series A）、《电子统计学期刊》（Electronic Journal of Statistics）、《心理计量学》（Psychometrika）。