机器学习与数据科学博士生系列论坛（第九十三期）—— Recursive Quantile Estimation: Non-Asymptotic Confidence Bounds
报告人：孙谌劼（北京大学）
时间：2025-10-30 16:00-17:00
地点：腾讯会议：331-2528-5257
摘要：Quantile estimation is fundamental in statistics, finance, and learning systems where interest lies in distributional characteristics beyond the mean. In modern large-scale or streaming data scenarios, traditional empirical estimators become computationally and memory inefficient, motivating recursive stochastic algorithms such as stochastic gradient descent (SGD) with Polyak–Ruppert averaging as scalable alternatives for online quantile estimation.
In this talk, we'll introduce non-asymptotic confidence bounds for the recursive quantile estimator based on SGD averaging. By bounding the moment generating function of the recursive estimate, the analysis yields exponentially decaying tail probability bounds under mild smoothness assumptions, improving upon previous algebraic-rate results. We'll also introduce its application in the problem of best arm identification in a multi-armed stochastic bandit setting under quantile preferences.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。
