|
Activity Number:
|
251
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, July 31, 2007 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Computing
|
| Abstract - #310191 |
|
Title:
|
Semisupervised Learning from Dissimilarity Data
|
|
Author(s):
|
Michael Trosset*+ and Carey E. Priebe
|
|
Companies:
|
Indiana University and Johns Hopkins University
|
|
Address:
|
PO Box 6424, Bloomington, IN, 47407,
|
|
Keywords:
|
embedding ; multidimensional scaling ; principal components ; classification ; discriminant coordinates
|
|
Abstract:
|
One approach to classifying objects for which only pairwise dissimilarities are available is to (1) embed the objects in Euclidean space (e.g., by classical multidimensional scaling), then (2) apply a conventional classification procedure (e.g., linear discriminant analysis). This two-stage approach is semisupervised: although only labeled objects can be used to construct the classifier, additional unlabeled objects can be used to facilitate construction of the representation space. We explore the extent to which a semisupervised approach improves on a fully supervised approach.
|
