|
Activity Number:
|
286
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, August 4, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Nonparametric Statistics
|
| Abstract - #305796 |
|
Title:
|
Applications of Spectral Kernel Methods
|
|
Author(s):
|
James Sharpnack*+
|
|
Companies:
|
Carnegie Mellon University
|
|
Address:
|
5506 Fifth Ave., Statistics Dept, Pittsburgh, PA, 15232,
|
|
Keywords:
|
diffusion processes ; spectral connectivity analysis ; information retrieval ; principal component analysis
|
|
Abstract:
|
Spectral connectivity analysis refers to a motley of methods relying on the spectral decomposition of kernels in order to preserve various notions of similarity between samples. We analyze the performance of principal component analysis, Laplacian eigenmaps and diffusion processes on a variety of data sets. We pay special attention to tuning parameters within these methods and assess the currently suggested heuristics. We find that many of these methods rely heavily on the original distance and the basis on which the data is first projected. We also explore the use of diffusion K-means for clustering samples in order to subsample the data set. These methods are applied to information retrieval problems, namely bag-of-words models, and image segmentation.
|
