摘要：This paper is concerned about a typical type of weak-supervision, the label noise problem.  A common setting for classification with label noise assumes that the noise level is independent of feature and known a priori. We consider the setting where a validation dataset with correct labels is available at learning time in addition to a large dataset with label noise. We argue that the classification with possibly feature-dependent noise in weakly-supervised settings can naturally be solved by a general logistic regression.  The rate-optimal estimators are obtained via maximizing a penalized joint likelihood function.  A sample-splitting-based method is further proposed for constructing confidence intervals for individual components of the regression vector, which enables us to identify label-noise-related features with error rate control.  The superiority of our method is demonstrated through asymptotic properties as well as numerical experiments.  A real example is also presented to illustrate how to use the proposed method in practice.

报告人简介：

王兆军，南开大学统计与数据科学学院执行院长/教授，国务院学位委员会统计学科评议组成员，全国统计教材编审委员会委员; 中国工业与应用数学学会副理事长， 中国统计教育学会副会长，中国工业统计教学研究会副会长，中国概率统计学会副理事长。曾任国家统计专家咨询委员会委员、中国现场统计研究会副理事长、天津市现场统计研究会理事长，天津工业与应用数学学会理事长，曾获国务院政府特贴、全国百篇优博指导教师、教育部自然科学二等奖及天津市自然科学一等奖。