主讲人：
Xiaojie Mao(Associate Professor in Management Science and Engineering at Tsinghua University)
主持老师：
（北大国发院）沈艳
参与老师：
（北大经院）王一鸣、王熙、刘蕴霆、王法、李少然、巩爱博
（北大国发院）黄卓、张俊妮
时间：
2026年5月8日（周五）
10：00-11:30
地点（线下）：
北京大学国家发展研究院承泽园131教室
主讲人简介：
Xiaojie Mao is an associate professor in Management Science and Engineering at Tsinghua University. He did his undergraduate in Mathematical Economics at Wuhan University and Ph.D. in Statistics and Data Science at Cornell University. His research interest lies in causal inference and data-driven decision-making. His research has appeared in top journals and conferences across multiple fields, such as Operations Research, Information Systems Research, Management Science, Journal of Machine Learning Research, Journal of the Royal Statistical Society Series B, NeurIPS, ICML, COLT, etc.
报告摘要：
The problem of selective label observations is common in many decision-making applications involving human subjects. In these applications, whether an individual’s outcome label can be observed depends on a certain decision. For example, in lending, the default status of a loan applicant cannot be observed if the loan is not approved, creating selection bias in labeled data. The selective label problem presents significant challenges for developing effective machine learning algorithms. In this talk, I will present our research on learning classifiers from selectively labeled data in both offline and online settings. In the offline setting, we follow the existing literature and consider selective labels arising from the decisions of multiple heterogeneous decision-makers. We formalize this setup through an instrumental variable framework, provide principled identification analysis, and propose a cost-sensitive learning algorithm to tackle the selective label problem. In the online setting, we additionally consider individuals’ strategic behaviors of manipulating their features to achieve favorable classification (e.g., get loan approval). We propose an online optimization algorithm to learn linear classifiers under both selective labels and strategic manipulation. We further prove that the expected regret of this algorithm is sublinear.
