Robust Nonnegative Matrix Factorization: An Approach to Dimension Reduction, Image Compression and A

发布日期: 2016-12-21  作者:    浏览次数: 424 


报告人:Sun Jiayang, Case Western Reserve University, Cleveland, USA

题目:Robust Nonnegative Matrix Factorization: An Approach to Dimension Reduction, Image Compression and Applications

时间:20161222 (周四) 下午330 - 430





We present a new approach, robust non-negative factorization (rNMF) for analyzing large and corrupted data with both fully automatic and semi-supervised controls. It applies a penalized criterion to strategically incorporate a trimming component to simultaneously handle a variety of noises and control the sparsity of the decomposition.  The resulting automatic rNMF works well when compared to popular NMF algorithms commonly used in practice and an existing robust NMF procedure,  as well as a robust PCA and a robust Singular Value Decomposition (SVD). As a sibling to the rNMF, our semi-supervised rNMF, called multistage rNMF (MrNMF) employs a hidden multi-level scheme to address further difficult areas that an unsupervised algorithm cannot properly handle.  We show the convergence of rNMF and demonstrate the power of rNMF and MrNMF for data cleaning, compression, and feature selection, using simulated tumor image data and images of faces subject to different types of corruptions. This is joint work with Y Ethan Xu, K. Lopiano and S. Young.



Professor at Department of Epidemiology and Biostatistics , Director of Center for Statistical Research, Computing and Collaboration (SR2c),  Director of Modern Biostatistics Concentration in the Ph.D. program of Epidemiology & Biostatistics. Elected Fellow of American Statistical Association (ASA) , Elected Fellow of Institute of Mathematical Statistics (IMS) , Elected Member of International Statistical Institute (ISI) .



  版权所有2009 ©  请勿转载和建立镜像© © 违者依法必究© © 上海师范大学数理学院