|
Activity Number:
|
436
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Learning and Data Mining
|
| Abstract - #304432 |
|
Title:
|
Clustering via Data Spectroscopy
|
|
Author(s):
|
Jared Schuetter*+ and Tao Shi
|
|
Companies:
|
The Ohio State University and The Ohio State University
|
|
Address:
|
Cockins Hall Room 404, Columbus, OH, 43210,
|
|
Keywords:
|
Spectral Clustering ; Data Mining ; Unsupervised Learning
|
|
Abstract:
|
Data clustering is often done using an affinity matrix containing kernel functions of pairwise distances. Spectral clustering algorithms use eigenvectors of this affinity matrix - or a function of it---to identify cluster labels of the points. A competitive spectral algorithm, Data Spectroscopy, exploits a no sign change property to ?nd eigenvectors of the affinity matrix representing each of the data clusters. Group labels are assigned to the points by comparing these vectors, and can be extended to any point in the domain of the data set. Advantages include automatic selection of the number of groups, detection of lower-dimensional manifolds, and robustness to relative group size. A sampling procedure has also been developed which allows approximation of the method in larger data sets for which the affinity matrix cannot be stored in memory.
|
