统计与数据科学系系列学术报告之五百一十一期

时    间：2026年5月21日（星期四）16:00-17:00

主持人：复旦大学 管理学院 统计与数据科学系 章玮 助理教授

地    点：史带楼503室

报告人：任明旸 副教授

              上海交通大学

题  目：Beyond parameter and overall transfer: some advances on transfer learning for causal graph learning and classification with label noise

摘   要：Statistical transfer learning has emerged as a hot topic in modern data science, enabling knowledge borrowing from related auxiliary domains to improve performance in a target domain. However, numerous existing methods rely heavily on overall parameter similarity, i.e., the auxiliary and target domains are assumed to share almost entirely identical parameters. Such assumptions may fail in many important real-world problems where only local structural correspondences exist or where the data generation processes are heterogeneous and model-free. In our recent works, we have gone far beyond parameter‑based overall transfer by addressing two challenging scenarios: causal graph reconstruction with only node‑level local similarities, and classification with label noise where no parametric model or overall similarity can be assumed. In this talk, I will first introduce the problem of reconstructing directed acyclic graphs (DAGs) and propose a novel set of structural similarity measures for different DAGs. A transfer DAG learning framework that effectively exploits auxiliary DAGs at the node level, along with its theoretical analysis, will be presented. Second, I will tackle classification with label noise under a common yet challenging data paradigm: a large coarse‑labeled dataset accompanied by a small expert‑verified clean dataset, and develop a novel transfer learning framework that is both model-agnostic and similarity-friendly, which is supported by rigorous statistical theory. Extensive experiments on synthetic data and real data analyses, including multi-source brain functional connectivity analysis, multi-source cancer gene regulatory networks, and pneumonia diagnosis from medical images, have validated the advantages of these methods.

个人简介：任明旸，上海交通大学数学科学学院长聘教轨副教授、博士生导师。博士毕业于中国科学院大学，期间于耶鲁大学生物统计系联合培养。曾于香港中文大学统计学系从事博士后研究。研究方向为异质数据分析、图模型和生物统计。以第一或通讯作者在JMLR, AOAS, Biometrics, Bioinformatics等期刊发表论文十余篇。主持和参与多项国家自然科学基金，入选上海市海外高层次人才计划和教育部海外博士后引才计划。曾获ICSA Junior Researcher Award Honorable Mention、International Biometric Society ENAR Distinguished Student Paper Award、中国科学院院长奖等奖项。

 统计与数据科学系

2026-4-30