Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.adass.org/adass/proceedings/adass02/reprints/P4-10.pdf Дата изменения: Wed Mar 12 01:46:51 2003 Дата индексирования: Tue Oct 2 11:28:48 2012 Кодировка: Поисковые слова: aircraft

Astronomical Data Analysis Software and Systems XII ASP Conference Series, Vol. 295, 2003 H. E. Payne, R. I. Jedrzejewski, and R. N. Hook, eds.

Automated Ob ject Classification with ClassX
A. A. Suchkov, R. J. Hanisch, R. L. White, M. Postman, & M. E. Donahue Space Telescope Science Institute T. A. McGlynn, L. Angelini, M.F. Corcoran, S.A. Drake, W.D. Pence, N. White, & E.L. Winter Goddard Space Flight Center F. Genova, F. Ochsenbein, P. Fernique, & S. Derriere Centre de Donnґes astronomiques de Strasbourg e Abstract. ClassX is a pro ject aimed at creating an automated system to classify X-ray sources and is envisaged as a prototype of the Virtual Observatory. As a system, ClassX creates a pipeline by integrating a network of classifiers with an engine that searches and retrieves multi-wavelength counterparts for a given target from the worldwide data storage media. At the start of the pro ject we identified a number of issues that needed to be addressed to make the implementation of such a system possible. The most fundamental are: (a) classification methods and algorithms, (b) selection and definition of classes (ob ject types), and (c) identification of source counterparts across multi-wavelength data. Their relevance to the pro ject ob jectives will be seen in the results below as we discuss ClassX classifiers.

1.

Classifiers

We apply machine learning methods to generate classifiers from `training' data sets, each set being a particular sample of ob jects with pre-assigned class names that have measured X-ray fluxes and, wherever possible, data from other wavelength bands. In this paper, a classifier is represented by a set of oblique decision trees (DT) induced by a DT generation system OC1. An X-ray source is input into a classifier as a set of X-ray fluxes and possibly data from the optical, infrared, radio, etc. The discussion below includes some results obtained with classifiers trained on the data from the ROSAT WGA, GSC2, and 2MASS catalogs. 1.1. Classifier Metrics

In order to quantify the quality and efficiency of classifiers, we have introduced a variety of metrics. They include the classifier's preference, Pij , which is the probability that a class i ob ject will be classified as class j (Figure 1); its affinity, 419 c Copyright 2003 Astronomical Society of the Pacific. All rights reserved.


420

Suchkov et al.

Figure 1. Preference, Pij , (left) and affinity, Aij , (right) for three classifiers trained using (a) X-ray magnitudes, (b) GSC2 magnitudes, and (c) both GSC2 and X-ray magnitudes along with coordinates and GSC2 "extended" vs. "point" source parameter. Notice that the OC1 classifiers separate stellar ob jects from non-stellar ones quite reliably. At the same time, a confusion between different types of stars or, say, QSO and AGN should be expected because of original misclassification and significant overlap of the respective ob ject types in the parameter space. Aij , which is the probability that an ob ject classified to class j (Figure 1), and the power, Si , which is that an ob ject classified as class i is indeed class randomly selected ob ject belongs to class i. Addi are completeness, Ci = Pii , and reliability, Ri = Aii 1.2. Classifier Networks as class i does in fact belong the ratio of the probability i to the probability that a tional useful characteristics .

Using different sets of training parameters (attributes), we get different classifiers for the same list of class names (e.g., Figure 1). We integrate them into a network, in which each classifier makes its own class assignment and is optimized for handling different tasks or different ob ject types. We envision that, having a set of X-ray sources, a user would generally select a certain classifier to make, for instance, the most complete list of candidate QSOs, but a different


Automated Ob ject Classification with ClassX

421

classifier would be used to make a most reliable list of such candidates. Additional classifiers would be selected to make similar lists for other ob ject types. Figure 1 suggests that one would prefer the xray gsc2 classifier to pick up cluster candidates, while AGNs call for the xray only classifier. 2. Training Set Deficiencies

A classifier is adversely impacted by source misclassification, counterpart misidentification, data bias, etc. As the training data improve, so do the classifiers. In Figure 2, about 50% of class "Stars" sources (stars without spectral classification) come from the LMC/SMC region. This introduces a coordinate bias that affects classifiers generated from those data. Certain metrics of a classifier can be improved if stars from the LMC/SMC region are dropped from the respective training sets. 3. Validating Pre-assigned Classes with ClassX

An X-ray source in a training set may have an inappropriate class name or incorrect optical or other counterpart. Candidates with these deficiencies can be identified as a classifier is applied to the training data. In Figure 2, an OC1 classifier is seen to noticeably enhance the contrast between "extended" and "point" sources for StarsKM, QSOs, and Clusters, suggesting that the sources contributingto that enhancement were probably misclassified in the trainingset. They can further be examined and then reclassified if warranted, which would improve the training set itself. 4. Counterpart Search Strategies with ClassX if a couninstance, ec. Thus, to search

Classifiers trained using optical counterparts proved to be much better terpart is selected as a brightest ob jects within 30 arcsec as opposed, for to a nearest ob ject within 30 arcsec or a brightest ob ject within 60 arcs classifier validation in ClassX offers a way to find the best strategies for multi-wavelength counterparts. 5. Class Ambiguity in ClassX

A class is rarely a clear-cut notion. One person's QSO is another's AGN or galaxy. The overlap of ob ject properties that often results in confusion in the ob ject type assignement is an essential issue for any classification system. With ClassX, one can isolate sources with a greater degree of class name ambiguity and look into why their classification in the training set differs from the OC1 classification (see Figure 2).


422

Suchkov et al.

Figure 2. X-ray soft versus hard magnitudes for classes from the WGA catalog (left) and the OC1 classifier (right). On the right, only the WGA class QSO is shown. Most of the brightest and the faintest WGA QSOs have been classified by the OC1 classifier as AGN and Cluster, respectively, which partly reflects the class name ambiguity for these sources. 6. ClassX Outputs name and the probability that the source outputs the probabilities that the source class name list. This is useful, for instance, ciation is with various classes in parameter

A network classifier outputs the class belongs to the assigned class. It also belongs, in fact, to other classes in the for assessinghow close the source asso space.