【报告摘要】Detecting structural breaks in high-dimensional, highly flexible models confronts two core challenges: 1) Computational burden: re-fitting a sophisticated learner for every candidate segment is prohibitive; 2) Estimation bias: scoring flexible learners with in-sample loss produces systematically optimistic estimates that misplace changepoints. This talk introduces a dual remedy. Reliever tiles the search space with a near-linear family of multiscale “relief” intervals, enabling any optimal or greedy algorithm to recycle a single model fit across many segments. The resulting O(n) total fitting cost still achieves (near) rate-optimal recovery of changepoints in high-dimensional regression and other rich settings. Cross-fitted evaluation partitions the data into folds, scoring each fold with models trained without it to obtain unbiased out-of-sample losses. Minimizing this criterion restores consistency (and rate-optimality) under mild predictive-accuracy assumptions. The strategies of Reliever and cross-fitting together form a principled, practical toolkit that renders changepoint analysis both fast and trustworthy, paving the way for its routine use in modern machine-learning pipelines.

【报告人简介】邹长亮，南开大学统计与数据科学学院教授。主要从事统计学及其与数据科学领域的交叉研究和实际应用。研究兴趣包括：预测性推断、变点和异常点检测、高维数据统计学习等。近年来在统计学和机器学习领域的权威期刊和会议上发表发表论文数十篇，入选爱思唯尔“中国高被引学者”。主持基金委优青、杰青、重点项目、重大项目课题和科技部重点研发计划课题等。任教育部科技委委员、全国应用统计专业硕士教学指导委员会委员、中国现场统计研究会副理事长等。