Large surveys producing tera- and petabyte-scale databases require machine-learning and knowledge discovery methods to deal with the overwhelming quantity of data and the difficulties of extracting concise, meaningful information with reliable assessment of its uncertainty. This study investigates the potential of a few machine-learning methods for the automated analysis of eclipsing binaries in the data of such surveys. We aim to aid the extraction of samples of eclipsing binaries from such databases and to provide basic information about the objects. We estimate class labels according to two classification systems, one based on the light curve morphology (EA/EB/EW classes) and the other based on the physical characteristics of the binary system (system morphology classes; detached through overcontact systems). Furthermore, we explore low-dimensional surfaces along which the light curves of eclipsing binaries are concentrated, to use in the characterization of the binary systems and in the exploration of biases of the full unknown Gaia data with respect to the training sets. We explore the performance of principal component analysis (PCA), linear discriminant analysis (LDA), random forest classification and self-organizing maps (SOM). We pre-process the photometric time series by combining a double Gaussian profile fit and a smoothing spline, in order to de-noise and interpolate the observed light curves. We achieve further denoising, and selected the most important variability elements from the light curves using PCA. We perform supervised classification using random forest and LDA based on the PC decomposition, while SOM gives a continuous 2-dimensional manifold of the light curves arranged by a few important features. We estimate the uncertainty of the supervised methods due to the specific finite training set using ensembles of models constructed on randomized training sets.
M. Suveges, F. Barblan, I. Lecoeur-Taibi, et. al.
Wed, 22 Feb 17
Comments: 20 pages, 22 figures. Accepted for publication in A&A