Sparse Discriminant Analysis and Multi-Collinearity with Applications to Image Analysis

• Activity Number: 476
• Type: Invited
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Line H. Clemmensen*+
• Companies: Technical University of Denmark
• Keywords: sparse discriminant analysis ; image analysis ; high dimensionality ; classification ; covariance assumptions ; multiple classes

Abstract

In an image analysis context data often have high correlations due to spatial and spectral relations between pixels. The high correlations are inherited to features when feature extraction is applied to the images. Here, we are concerned with the classification setting, and in particular extending discriminant analysis to the high-dimensional case (p>n). The sparse discriminant analysis based on optimal scoring was developed for this setting. We examine the assumptions of independence versus collinearity made for sparse discriminant analysis as well as a variety of related methods, and the relation to the data of interest. Furthermore, we discuss classification of multiple classes as opposed to binary classification. Examples from image analysis will be given.