统计与数据科学系系列学术报告之五百二十九期

时间：2026年11月18日（星期三）10:00-11:00

主持人：复旦大学管理学院统计与数据科学系沈娟副教授

地点：史带楼302室

报告人：Jingfei Zhang 张菁菲教授

Emory University

题目：Statistical Inference with AI-Generated Predictive Distributions

摘要：In many applications, labeled data are expensive to obtain, while unlabeled data and predictions from pretrained models such as LLMs are readily available. Prediction-powered inference (PPI) leverages such predictions to improve the statistical efficiency with limited labeled data. However, existing PPI methods typically rely on point predictions, even when richer predictive distributions are available.

In this talk, I will introduce distribution-informed prediction-powered inference (DiPPI), a general framework that uses predictive distributions as auxiliary information to further improve statistical efficiency. We characterize the optimal use of this information through score calibration, derive the oracle efficiency for finite-dimensional representations of predictive distributions, and establish theoretical guarantees for positive learning with a cross-fitted estimator.

个人简介：Dr. Zhang is the Goizueta Foundation Term Chair Professor of Information Systems & Operations Management (ISOM) at Emory University’s Goizueta Business School, where she currently serves as area chair for ISOM. Her research focuses on the statistical analysis of networks, graphs and tensors, with applications in business and biomedical sciences. She is an elected member of the International Statistical Institute (ISI) and a Fellow of the American Statistical Association (ASA). She currently serves as an Associate Editor for the Journal of the American Statistical Association, Annals of Applied Statistics and Journal of the Royal Statistical Society Series B.