题目：Discovering Relevant Variables Without a Model: Generative Conditional Independence Testing with False Discovery Control

主讲人：华盛顿大学圣路易斯分校统计与数据科学系 邵晓峰教授

主持人：西南财经大学统计与数据科学学院 常晋源教授

时间：2026年6月23日（周二）上午10:30

地点：西南财经大学光华校区光华楼1003会议室

报告摘要：

High-dimensional datasets often contain many potential predictors, but only a small subset may be truly relevant to the response. Classical variable selection methods can work well when their modeling assumptions are correct, but they may miss signals that are nonlinear, interaction-driven, or hidden in the full conditional distribution rather than in the conditional mean. This talk presents a nonparametric variable selection framework that treats relevance as a conditional independence question: a variable is selected only if it still carries information about the response after accounting for all other variables.    

The proposed method uses generative neural networks to learn the conditional distributions needed to create synthetic “null” samples, then compares the observed data with these null samples through a kernel-based discrepancy measure. A sample-splitting and studentization strategy yields a simple normal calibration, avoiding computationally expensive bootstrap or permutation procedures. By combining the resulting coordinate-wise tests with a multiple-testing thresholding rule, the method provides asymptotic control of the false discovery rate in high-dimensional settings.    

The talk will highlight the main statistical ideas, the role of generative modeling, and why the method is especially useful for detecting nonlinear and interaction effects that may be missed by existing model-free procedures based on fixed transformations of the response. Simulations show strong performance in nonlinear settings while remaining competitive in linear cases, and an application to cancer drug-response data identifies biologically meaningful mutations related to BRAF-targeted treatment.    

主讲人简介：

Xiaofeng Shao joined Washington University in January 2025 as a Professor of Statistics & Data Science, with a joint appointment in the Department of Economics. Prior to this, he served on the Statistics faculty at the University of Illinois at Urbana-Champaign for 18 years. He is a Fellow of the Institute of Mathematical Statistics and the American Statistical Association.    

His primary research interests include econometrics, time series analysis, change-point analysis, high-dimensional statistics, nonparametric statistics, and functional data analysis. Recently, his work has focused on developing models and methods for analyzing object-valued time series and creating innovative inference procedures for high-dimensional and imaging data, leveraging advancements in machine learning.