DIVERGENCE BASED FEATURE-SELECTION FOR MULTIMODAL CLASS DENSITIES

Authors
NOVOVICOVA J PUDIL P KITTLER J
Citation
J. Novovicova et al., DIVERGENCE BASED FEATURE-SELECTION FOR MULTIMODAL CLASS DENSITIES, IEEE transactions on pattern analysis and machine intelligence, 18(2), 1996, pp. 218-223
Citations number
11
Categorie Soggetti
Computer Sciences","Computer Science Artificial Intelligence","Engineering, Eletrical & Electronic
Journal title
IEEE transactions on pattern analysis and machine intelligence → ACNP
ISSN journal
01628828
Volume
18
Issue
2
Year of publication
1996
Pages
218 - 223
Database
ISI
SICI code
0162-8828(1996)18:2<218:DBFFMC>2.0.ZU;2-I
Abstract
A new feature selection procedure based on the Kullback J-divergence b etween two class conditional density functions approximated by a finit e mixture of parameterized densities of a special type is presented. T his procedure is suitable especially for multimodal data. Apart from f inding a feature subset of any cardinality without involving any searc h procedure, it also simultaneously yields a pseudo-Bayes decision rul e. Its performance is tested on real data.