The Multicategory Support Vector Machine

[Grace Wahba]

I'm pleased to announce
 "Classification of Multiple Cancer Types by Multicategory Support 
       Vector Machines Using Gene Expression Data" 
                     by 
 Yoonkyung Lee and Cheol-Koo Lee, UW-Madison Statistics Dept TR 1051 (2002).
 This paper (and related papers) are accessible 
 via the home pages:
 http://www.stat.wisc.edu/~yklee  
              or 
 http://www.stat.wisc.edu/~wahba
            
       Abstract:

Monitoring gene expression profiles is a novel approach in cancer 
diagnosis. Several studies showed that prediction of cancer types 
using gene expression data is promising and very informative.
The Support Vector Machine (SVM) is one of the classification methods
successfully applied to the cancer diagnosis problems using gene expression 
data. However, its extension to more than two classes was not 
obvious, which might impose limitations in its application to multiple tumor 
types. In this paper, we analyze a couple of published multiple cancer types 
data sets by the multicategory SVM, which is a recently proposed extension of
the binary SVM.    
......................
.......................
The Multicategory Support Vector Machine was proposed in 
 "Multicategory Support Vector Machines", Yoonkyung Lee, Yi Lin and Grace Wahba
 UW-Madison Statistics Dept TR 1043 (2001), also accessible as above. 
.........................
Multicategory `Soft' classification was proposed in 
 "Smoothing Spline Analysis of Variance for Polychotomous Response Data"
 Xiwu Lin, UW-Madison Statistics Dept TR 1003 (1998), available via the wahba home pg. 
 It can be argued that the SVM can be considered a `hard' classifier and 
 while the penalized likelihood estimate (as in TR 1003)is a `soft' 
 classifier, and it can be argued that the
 the SVM is more appropriate where the attribute 
 data is relatively sparse and the category overlap is relatively 
 small while the soft classifier may be more appropriate where the 
 attribute data is relatively dense and the category overlap 
 is more substantial.