ON REDUCING DIMENSIONALITY OF DISSIMILARITY MATRICES FOR OPTIMIZING DBC - An Experimental Comparison

Sang-Woon Kim

2010

Abstract

One problem of dissimilarity-based classifications (DBCs) is the high dimensionality of dissimilarity matrices. To address this problem, two kinds of solutions have been proposed in the literature: prototype selection (PS) based methods and dimensionality reduction (DR) based methods. The DR-based method consists of building the dissimilarity matrices using all the available training samples and subsequently applying some of the standard DR schemes. On the other hand, the PS-based method works by directly choosing a small set of representatives from the training samples. Although DR-based and PS-based methods have been explored separately by many researchers, not much analysis has been done on the study of comparing the two. Therefore, this paper aims to find a suitable method for optimizing DBCs by a comparative study. In the experiments, four DR and four PS methods are used to reduce the dimensionality of the dissimilarity matrices, and classification accuracies of the resultant DBCs trained with two real-life benchmark databases are analyzed. Our empirical evaluation on the two approaches demonstrates that the DR-based method can improve the classification accuracies more than the PS-based method. Especially, the experimental results show that the DR-based method is clearly more useful for nonparametric classifiers, but not for parametric ones.

Paper Citation

in Harvard Style

Kim S. (2010). ON REDUCING DIMENSIONALITY OF DISSIMILARITY MATRICES FOR OPTIMIZING DBC - An Experimental Comparison . In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Volume 1: ICAART, ISBN 978-989-674-021-4, pages 235-240. DOI: 10.5220/0002713002350240

in Bibtex Style

@conference{icaart10,
author={Sang-Woon Kim},
title={ON REDUCING DIMENSIONALITY OF DISSIMILARITY MATRICES FOR OPTIMIZING DBC - An Experimental Comparison},
booktitle={Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Volume 1: ICAART,},
year={2010},
pages={235-240},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0002713002350240},
isbn={978-989-674-021-4},
}

in EndNote Style

TY - CONF
JO - Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Volume 1: ICAART,
TI - ON REDUCING DIMENSIONALITY OF DISSIMILARITY MATRICES FOR OPTIMIZING DBC - An Experimental Comparison
SN - 978-989-674-021-4
AU - Kim S.
PY - 2010
SP - 235
EP - 240
DO - 10.5220/0002713002350240