报告时间：2023年6月6日

               上午10：00

报告地点：腾讯会议ID：149 937 008

报告嘉宾：王军辉

报告主题：Chain graph models: identifiability, estimation and asymptotics

报告摘要

Chain graph models: identifiability, estimation and asymptotics

In this talk, we consider a flexible chain graph (CG) model, which admits both undirected and directed edges in one graph and thus can encode much more diverse relations among objects. We first establish the identifiability conditions for the CG model through a low rank plus sparse matrix decomposition, where the sparse matrix implies the sparse undirected edges within each chain component and the low rank matrix implies the presence of hub nodes with multiple children or parents. On this ground, we develop an efficient estimation method for reconstructing the CG structure, which first identifies the chain components via estimated undirected edges, determines the causal ordering of the chain components, and eventually estimates the directed edges among the chain components. Its theoretical properties will be discussed in terms of both asymptotic and finite-sample probability bounds on model estimation and graph reconstruction. The advantage of the proposed method is also demonstrated through extensive numerical experiments on both synthetic data and the Standard & Poor’s 500 index data.

作者简介

王军辉教授现为香港中文大学统计系教授。他本科毕业于北京大学，研究生毕业于美国明尼苏达大学并获得统计学博士学位。他的研究方向包括统计机器学习及其在生物医学、经济、金融、和信息技术上的应用。他的研究成果广泛发表于JASA,Biometrika,JMLR和NeurIPS等统计及机器学习的顶级期刊和会议，并担任JASA，AOAS,StatisticaSinica等主流期刊的副主编。