Augmented embedding of dissimilarity data into (pseudo-)Euclidean spaces

None, None; None, None; None, None; None, None

Augmented embedding of dissimilarity data into (pseudo-)Euclidean spaces

Conference Paper (2006)

Author(s)

A Harol (TU Delft - Multimedia Computing)

EM Pekalska (TU Delft - Multimedia Computing)

S Verzakov (TU Delft - Multimedia Computing)

RPW Duin (TU Delft - Multimedia Computing)

CWTS 0.75 <= JFIS < 2.00 Conference contrib. refereed

To reference this document use

https://resolver.tudelft.nl/uuid:3f240841-f913-4035-974e-7dacf5e4d509

More Info

expand_more

Publication Year

2006

Pages (from-to)

613-621

Publisher

Springer

Event

Joint IAPR International Workshops SSPR 2006 and SPR 2006, Hong Kong, China (2006-08-17 - 2006-08-19), Heidelberg

Downloads counter

93

Abstract

Pairwiseproximitiesdescribethepropertiesofobjectsintermsoftheirsimilarities.Byusingdi¿erentdistance-basedfunctionsonemayencodedi¿erentcharacteristicsofagivenproblem.However,tousetheframeworkofstatisticalpatternrecognitionsomevectorrepresentationshouldbeconstructed.Oneofthesimplestwaystodothatistode¿neanisometricembeddingtosomevectorspace.Inthiswork,wewillfocusonalinearembeddingintoa(pseudo-)Euclideanspace.
Thisisusuallywellde¿nedfortrainingdata.Someinadequacy,however,appearswhenprojectingnewortestobjectsduetotheresultingprojectionerrors.Inthispaperweproposeanaugmentedembeddingalgorithmthatenlargesthedimensionalityofthespacesuchthattheresultingprojectionerrorvanishes.Ourpreliminaryresultsshowthatitmayleadtoabetterclassi¿cationaccuracy,especiallyfordatawithhighintrinsicdimensionality.

No files available

Metadata only record. There are no files for this record.