A machine learning approach to Cepheid variable star classification using data alignment and maximum likelihood Academic Article uri icon

abstract

  • Our study centers on the classification of two subtypes of Cepheid variable stars. Such a classification is relatively easy to obtain for nearby galaxies, but as we incorporate new galaxies, the cost of labeling stars calls for some form of model adaptation. Adapting a predictive model to differentiate Cepheids across galaxies is difficult because of the sample bias problem in star distribution (due to the limitation of telescopes in observing faint stars as we try to reach distant galaxies). In addition, estimating the luminosity of a star as we reach distant galaxies carries some inevitable shift in the data distribution. We propose an approach to predict the class of Cepheid stars on a target domain, by first building a model on an "anchor" source domain. Our methodology then shifts the target data until it is well aligned with the source data by maximizing two different likelihood functions. Experimental results with two galaxy datasets (Large Magellanic Cloud as the source domain, and M33 as the target domain), show the efficacy of the proposed method. 2013 Elsevier B.V.

published proceedings

  • ASTRONOMY AND COMPUTING

author list (cited authors)

  • Vilalta, R., Gupta, K. D., & Macri, L.

citation count

  • 11

complete list of authors

  • Vilalta, Ricardo||Gupta, Kinjal Dhar||Macri, Lucas

publication date

  • August 2013