JAIST Repository >
School of Knowledge Science >
JAIST Research Reports >
Research Report - School of Knowledge Science: ISSN 1347-1570 >
KS-RR-2009 >
Please use this identifier to cite or link to this item:
http://hdl.handle.net/10119/8448
|
Title: | Feature similarity: geometrical framework and discriminative kernel learning |
Authors: | Nguyen, Canh Hao Ho, Tu Bao |
Issue Date: | 2009-02-20 |
Publisher: | 北陸先端科学技術大学院大学知識科学研究科 |
Magazine name: | Research report (School of Knowledge Science, Japan Advanced Institute of Science and Technology) |
Volume: | KS-RR-2009-001 |
Start page: | 1 |
End page: | 22 |
Abstract: | We propose geometrical models of features of a learning problem. We show that there are two alternative ways to interpret similarities between features. The syntactic way is well studied and applied in feature selection while the semantic way, in the context of kernel methods, is less understood. We show that the latter is equivalent to feature semantic similarity (FSS) and there are a number of methods that fall into this framework. We analyze to show relations among these methods and differences to feature selection ones. Our analysis shows a natural extension to all these methods. It automatically suggests that this framework can be applied in a general context. We also note that all the methods using FSS are inherently unsupervised in nature. None of these methods make use of labels for classification or regression tasks. On the other hand, the feature selection counterparts consider labels as an important information. Therefore, we propose an algorithm to learn the feature proximity matrix for FSS for supervised tasks. We show the merit of our algorithm in various applications. |
URI: | http://hdl.handle.net/10119/8448 |
Material Type: | publisher |
Appears in Collections: | KS-RR-2009
|
Files in This Item:
File |
Description |
Size | Format |
KS-RR-2009-001.pdf | | 31474Kb | Adobe PDF | View/Open |
|
All items in DSpace are protected by copyright, with all rights reserved.
|