NASA Logo

NTRS

NTRS - NASA Technical Reports Server

Back to Results
Multiclass Reduced-Set Support Vector MachinesThere are well-established methods for reducing the number of support vectors in a trained binary support vector machine, often with minimal impact on accuracy. We show how reduced-set methods can be applied to multiclass SVMs made up of several binary SVMs, with significantly better results than reducing each binary SVM independently. Our approach is based on Burges' approach that constructs each reduced-set vector as the pre-image of a vector in kernel space, but we extend this by recomputing the SVM weights and bias optimally using the original SVM objective function. This leads to greater accuracy for a binary reduced-set SVM, and also allows vectors to be 'shared' between multiple binary SVMs for greater multiclass accuracy with fewer reduced-set vectors. We also propose computing pre-images using differential evolution, which we have found to be more robust than gradient descent alone. We show experimental results on a variety of problems and find that this new approach is consistently better than previous multiclass reduced-set methods, sometimes with a dramatic difference.
Document ID
20080021359
Acquisition Source
Jet Propulsion Laboratory
Document Type
Preprint (Draft being sent to journal)
External Source(s)
Authors
Tang, Benyang
(Jet Propulsion Lab., California Inst. of Tech. Wrightwood, CA, United States)
Mazzoni, Dominic
(Jet Propulsion Lab., California Inst. of Tech. Pasadena, CA, United States)
Date Acquired
August 24, 2013
Publication Date
July 25, 2006
Subject Category
Mathematical And Computer Sciences (General)
Meeting Information
Meeting: 23rd International Conference on Machine Learning
Location: 23rd International Conference on Machine Learning
Country: United States
Start Date: June 25, 2006
End Date: June 29, 2006
Distribution Limits
Public
Copyright
Other
Keywords
reduced set methods
Kernel pre-image
multiclass
differential evolution
support vector machines

Available Downloads

There are no available downloads for this record.
No Preview Available