|
Activity Number:
|
270
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #305204 |
|
Title:
|
Choosing a Dissimilarity Representation for Pattern Classification (and How to Use This Choice to Improve Your Results)
|
|
Author(s):
|
Adam Cardinal-Stakenas*+ and Carey E. Priebe and Zhiliang Ma and Youngser Park and Jeffrey L. Solka
|
|
Companies:
|
Johns Hopkins University and Johns Hopkins University and Johns Hopkins University and Johns Hopkins University and Naval Surface Warfare Center
|
|
Address:
|
6720 Second Morning Ct, Columbia, MD, 21045,
|
|
Keywords:
|
classification ; dissimilarity represenation ; multidimensional scaling
|
|
Abstract:
|
The dissimilarity representation is well understood to be a useful technique for exploiting data to solve pattern classification problems. In particular, alone or in concert with multidimensional scaling, dissimilarities feature prominently in many useful regularization techniques. For most problems, many appropriate dissimilarities exist. Hence, we must answer the question of how to choose the dissimilarity most well suited for the problem at hand. In this talk we will discuss a variety of techniques for choosing a dissimilarity measure and demonstrate how these techniques can be used on a variety of statistical inference problems. In particular we present results on how our methodologies can be applied when one wishes to effectively combine dissimilarity representations to improve classifier performance.
|
