AAS 200th meeting, Albuquerque, NM, June 2002
Session 60. Building a Virtual Observatory
Display, Wednesday, June 5, 2002, 10:00am-7:00pm, SW Exhibit Hall

[Previous] | [Session 60] | [Next]

[60.03] Classifying X-Ray Sources Using Multi-Wavelength Data

T.A. McGlynn (NASA/GSFC), A. A. Suchkov (STScI), L. Angelini, M.F. Corcoran, S.A. Drake, W.D. Pence, E.L. Winter (NASA/GSFC), R.J. Hanisch, R.L. White, M. Postman, M.E. Donahue (STScI), F. Genova, F. Ochsenbein, P. Fernique, M. Wenger (CDS)

We present the results of an ongoing project to build a system for automated classification of X-ray sources that uses a trainable classifier (currently OC1). Training data include X-ray measurements as well as optical, radio, and infrared data for objects with known classification. The trained classifier can be applied to make class (object type) assignment for X-ray sources of unknown nature. We have identified a number of issues that need to be addressed within this approach:

(a) Given that most of the classes in any training set are represented by too few objects, what is the best way to combine objects into broader classes? (b) As X-ray sources are cross-correlated with optical and/or infrared sources, how can the most likely counterpart among multiple candidates within the source error circle be identified? (c) How do the classifier and classification results depend on source attributes selected for class differentiation? (d) Many sources look as distinct different classes in different wavelength bands. What is the most effective way of combining multi-wavelength data for such objects for classification purposes?

In this part of the project, we have conducted a series of experiments to answer these questions. They have allowed us, in particular, to compare algorithms for selection of optical and infrared counterparts for X-ray sources; these have been used to both build training data sets and complement the catalog of X-ray sources with optical/infrared data. The experiments have also been used to quantify the quality and efficiency of a classifier in terms of the classifier's metrics, such as completeness, reliability, power, preference, etc. We conclude that, in general, classification would require a network of classifiers rather a single classifier, each producing an optimal classification for a particular class or set of classes.

[Previous] | [Session 60] | [Next]

Bulletin of the American Astronomical Society, 34
© 2002. The American Astronomical Soceity.