报告题目：Optimal multiple-prediction-powered inference for convolution smoothed

quantile regression

报告人： 王磊 教授 （南开大学）

报告时间：2026年11月22 日（周日）上午9：30-10：30

报告地点：学院会议室106

报告摘要: This paper develops a prediction-powered inference (PPI) framework for quantile regression by integrating convolution smoothing techniques and post-prediction inference. Within this framework, we propose a principled, data-adaptive weighting hierarchy (general-weighting,

diagonal-weighting and scalar-weighting) to leverage single or multiple machine learning (ML)

predictions of unknown quality for valid statistical inference. Under mild regularity conditions, the

proposed PPI estimators are asymptotically normal without assumptions on the quality of the

predictions. A unified minimization criterion based on the trace of the asymptotic covariance matrix is proposed to derive the optimal weighting schemes. By plugging in consistently estimated optimal weights, the resulting PPI estimators enjoy both safety and adaptivity. To address the finite-sample instability of weighting matrix estimation, we further propose ridge-regularization and sparsityregularization methods. Extensive numerical experiments under both synthetic and highly realistic ML prediction settings, together with real-data applications, demonstrate the favorable performance of the proposed PPI estimators in terms of finite-sample efficiency and computational stability.

报告人简介：王磊，南开大学疯av 教授、博导、百名青年学科带头人。研究方向是统计学习和复杂数据分析，已在统计学期刊Biometrika、JMLR、IEEE TIT、AOAS、Bernoulli、JCGS、Statistica Sinica 等发表学术论文多篇，主持3 项国家自然科学基金和1项天津市自然科学基金项目。