Variable selection for binary classification of spectral data : the KDR algorithm /
Classification has been approached almost exclusively as an extension of classical mean separation procedures, which depend upon the law of large numbers for their validity. In classification, however, we are always in the realm of the single observation, and new procedures, which focus on control...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
2001.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=728909211&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | Classification has been approached almost exclusively as an extension of classical mean separation procedures, which depend upon the law of large numbers for their validity. In classification, however, we are always in the realm of the single observation, and new procedures, which focus on control of the variability of the classification rule, are called for. This paper discusses several alternative classification methods, including a likelihood ratio algorithm that performs as well as traditional Fisherian methods and two additional algorithms (a geometric mean separation maximization algorithm and a distance ratio algorithm) which demonstrate substantially superior robustness to the methods currently in use. |
|---|---|
| Item Description: | Vita. "Major Subject: Statistics". |
| Physical Description: | xviii, 316 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 67-78). |