Incorporating Domain Specific Information into Gaia Source Classification
Astronomy is in the age of large scale surveys in which the gathering of multidimensional data on thousands of millions of objects is now routine. Efficiently processing these data — classifying objects, searching for structure, fitting astrophysical models — is a significant conceptual (not to mention computational) challenge. While standard statistical methods, such as Bayesian clustering, k-nearest neighbours, neural networks and support vector machines, have been successfully applied to some areas of astronomy, it is often difficult to incorporate domain specific information into these. For example, in astronomy we often have good physical models for the objects (e.g. stars) we observe. That is, we can reasonably well predict the observables (typically, the stellar spectrum or colours) from the astrophysical parameters (APs) we want to infer (such as mass, age and chemical composition). This is the “forward model”: The task of classification or parameter estimation is then an inverse problem. In this paper, we discuss the particular problem of combining astrometric information, effectively a measure of the distance of the source, with spectroscopic information.
KeywordsSupport Vector Machine Dark Matter Probability Vector Proper Motion Stellar Population
Unable to display preview. Download preview PDF.
- CHANG, C.-C. and LIN C.-J. (2001) : Libsvm: A library for support vector machines. (Tech-nical report) URL http://www.csie.ntu.edu.tw/∼cjlin/libsvm
- VAPNIK, V. (1995) The Nature of Statistical Learning Theory. Springer Verlag, New York. WU T.-F. and WENG R.C., (2004): Probability estimates for multi-class classification by pair-wise coupling. Journal of Machine Learning Research, 5:975-1005Google Scholar