Speaker：Zhaojun Wang (Nankai University)

Time：2024-5-16, 16:00

Location：Conference Room 105 at Experiment Building at Haiyun Campus

Abstract:

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. 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.