Документ взят из кэша поисковой машины. Адрес
оригинального документа
: http://www.stecf.org/conferences/adass/adassVII/christliebn.html
Дата изменения: Mon Jun 12 18:51:51 2006 Дата индексирования: Tue Oct 2 00:52:07 2012 Кодировка: Поисковые слова: п п п п п п п п |
Next: Information Mining in Astronomical Literature with Tetralogie
Up: Astrostatistics and Databases
Previous: Noise Detection and Filtering using Multiresolution Transform Methods
Table of Contents -- Index -- PS reprint -- PDF reprint
N. Christlieb1 and L. Wisotzki
Hamburger Sternwarte, Gojenbergsweg 112, D-21029 Hamburg, Germany
G. Graßhoff
Institut für Wissenschaftsgeschichte, Georg-August-Universität
Göttingen
A. Nelke and A. Schlemminger
Philosophisches Seminar, Universität Hamburg
1E-mail: nchristlieb@hs.uni-hamburg.de
A few years ago we started to develop methods for the systematic exploitation of the stellar content of the survey by means of automatic spectral classification. A short overview of the scientific objectives is given in Christlieb et al. (1997) and Christlieb et al. (1998), where also a detailed description of the classification techniques can be found. In this paper we report on the software system LINNé, that has been designed for the development and evaluation of classification models.
A classification model (CM) consists of the following components:
The aim of LINNé is to permit easy and well controlled access to the variation of the model components and effective means to evaluate the resulting quality of classification. The performance of a model with classification aim (1) can be evaluated by e. g. the total number of misclassifications, estimated with the leaving-one-out method (Hand 1981); in case of aim (3) the model is usually assessed by the number of misclassifications between the target class and the other classes.
The core of LINNé was implemented in an object-oriented extension to
Prolog, with the numerical routines - e. g. for estimation of the parameters
of the multivariate normal distributions - written in C. To facilitate user
interaction and to ensure effective control over the model components and
performance, a graphical user interface (GUI) for LINNé was developed
(see Figure 1).
The results of the classification model evaluation are presented on the client. A confusion matrix and loss matrix window assists the user in the analysis of the model. The user may then alter components and repeat the evaluation to improve the model step by step.
The search for the optimal feature set can also be done automatically. Since the set of available features may easily become too large to perform exhaustive search among all possible combinations, apart from the exhaustive search a hill-climbing like, stepwise search has been implemented. It can be controlled from the client side, using different strategies and branching parameters.
LINNé also provides a tool for the systematic and efficient adjustment
of loss factors (see Figure 2).
Once a CM has been established, evaluated, and the evaluation has pleased
the user, its parameters can be exported to MIDAS tables. The
classification of spectra of unknown classes can then be carried out under
MIDAS. The typical computing time for the classification of all spectra on
one HES plate - mapping of the sky and yielding
typically
non-disturbed spectra with S/N>10 - is less
than 5 min on a Linux PC with a Pentium 133 MHz processor.
So far LINNé has been used for some first test applications, i. e.
compilation of a sample of extremely metal poor halo stars and a search for
FHB/A stars. It will be developed further and extended in functionality and
will be applied to the exploitation of the huge HES data base, which will
finally consist of digitised objective prism spectra.
N.C. acknowledges an accommodation grant by the conference organizers. This work was supported by the Deutsche Forschungsgemeinschaft under grants Re 353/40-1 and Gr 968/3-1.
Christlieb, N. et al. 1997, in Wide-Field Spectroscopy, ed. Kontizas, E. et al., Kluwer, Dordrecht, 109
Christlieb, N. et al. 1998, to appear in Data Highways and Information Flooding, a Challenge for Classification and Data Analysis, ed. Balderjahn, I. et al., Springer, Berlin.
Hand, D. 1981, Discrimination and Classification, Wiley & Sons, New York.
Wisotzki, L. et al. 1996, A&AS, 115, 227
Next: Information Mining in Astronomical Literature with Tetralogie
Up: Astrostatistics and Databases
Previous: Noise Detection and Filtering using Multiresolution Transform Methods
Table of Contents -- Index -- PS reprint -- PDF reprint