【专家简介】：唐炎林，华东师范大学统计学院教授，博士生导师，统计学系主任；入选国家青年高层次人才计划、上海市浦江人才计划。主要研究方向为分位数回归、共形预测、高维异质性数据统计推断，主持多项国家自然科学基金、上海市自然科学基金，担任SCI期刊Statistica Sinica、Journal of the Korean Statistical Society的编委。在Biometrika、JRSSB、PNAS、Biometrics等发表论文近50篇。
【报告摘要】：In many real-world applications, statistical inference is challenged by the scarcity of labeled outcomes due to high data collection costs and technical barriers.
In the motivating MIMIC-III dataset, we aim to evaluate in-hospital illness severity among publicly insured patients, yet outcome data for this group are limited.
To address this challenge, we propose a framework for constructing efficient conformal prediction sets for target outcomes by leveraging an auxiliary source distribution with abundant labeled data that is related to the target distribution through a target shift assumption.
When target labels are unavailable, prediction relies solely on source data; when partial labels are observed, they can be incorporated to improve predictive efficiency.
To overcome data non-exchangeability and distributional non-identifiability, we identify the likelihood ratio by aligning covariate distributions of the source and target domains within a finite B-spline function space.
We further develop a weight-adjusted conditional density estimator to construct highest predictive density sets that accommodate complex error structures, including asymmetry and multimodality.
This estimator models the source conditional density via a quantile process and applies weighting transformations to approximate the target conditional density.
We establish theoretical guarantees for the proposed method and demonstrate its finite-sample performance through simulation and an application to the MIMIC-III clinical database.
【报告时间】：2026年09月21日（周一）15:40-16:40
【报告地点】：位育楼417
