Presentation 2003/9/8
Feature subset selection using restriction kernels
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) This paper presents a new feature subset selection algorithm than can take into account higher order correlation between variables. The algorithm is a kind of wrapper methods using Support Vector Machines (SVMs) for learning classifiers represented as hyperplanes spanned by combinations of variables. It is known that kernel functions enable efficient learning of the high dimensional hyperplanes, while this paper considers another use of kernel functions for analyzing the learned classifiers to determine irrelevant variables. In the analysis, the algorithm computes the restriction of a classifier obtained by removing the components containing a variable, and the variable is identified as irrelevant if the restriction discriminates data as well as the classifier. Although there exist numerous components to be removed, it is shown that the restriction can be computed efficiently by using restriction kernels. It is also shown that the presented algorithm outperforms existing algorithms in empirical studies on the synthetic data sets. Furthermore, the algorithm is applied to text categorization tasks and an encouraging result is obtained.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) feature selection / support vector machine / kernel methods / text categorization
Paper # AI2003-46
Date of Issue

Conference Information
Committee AI
Conference Date 2003/9/8(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Vice Chair

Paper Information
Registration To Artificial Intelligence and Knowledge-Based Processing (AI)
Language ENG
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Feature subset selection using restriction kernels
Sub Title (in English)
Keyword(1) feature selection
Keyword(2) support vector machine
Keyword(3) kernel methods
Keyword(4) text categorization
1st Author's Name Ken SADOHARA
1st Author's Affiliation National Institute of Advanced Industrial Science and Technology (AIST)()
Date 2003/9/8
Paper # AI2003-46
Volume (vol) vol.103
Number (no) 305
Page pp.pp.-
#Pages 6
Date of Issue