The EPOCH (EROS-2 periodic variable star classification using machine learning) project aims to detect periodic variable stars in the EROS-2 light curve database. In this paper, we present the first result of the classification of periodic variable s
tars in the EROS-2 LMC database. To classify these variables, we first built a training set by compiling known variables in the Large Magellanic Cloud area from the OGLE and MACHO surveys. We crossmatched these variables with the EROS-2 sources and extracted 22 variability features from 28 392 light curves of the corresponding EROS-2 sources. We then used the random forest method to classify the EROS-2 sources in the training set. We designed the model to separate not only $delta$ Scuti stars, RR Lyraes, Cepheids, eclipsing binaries, and long-period variables, the superclasses, but also their subclasses, such as RRab, RRc, RRd, and RRe for RR Lyraes, and similarly for the other variable types. The model trained using only the superclasses shows 99% recall and precision, while the model trained on all subclasses shows 87% recall and precision. We applied the trained model to the entire EROS-2 LMC database, which contains about 29 million sources, and found 117 234 periodic variable candidates. Out of these 117 234 periodic variables, 55 285 have not been discovered by either OGLE or MACHO variability studies. This set comprises 1 906 $delta$ Scuti stars, 6 607 RR Lyraes, 638 Cepheids, 178 Type II Cepheids, 34 562 eclipsing binaries, and 11 394 long-period variables. A catalog of these EROS-2 LMC periodic variable stars will be available online at http://stardb.yonsei.ac.kr and at the CDS website (http://vizier.u-strasbg.fr/viz-bin/VizieR).
The optical light curves of many quasars show variations of tenths of a magnitude or more on time scales of months to years. This variation often cannot be described well by a simple deterministic model. We perform a Bayesian comparison of over 20 de
terministic and stochastic models on 6304 QSO light curves in SDSS Stripe 82. We include the damped random walk (or Ornstein-Uhlenbeck [OU] process), a particular type of stochastic model which recent studies have focused on. Further models we consider are single and double sinusoids, multiple OU processes, higher order continuous autoregressive processes, and composite models. We find that only 29 out of 6304 QSO lightcurves are described significantly better by a deterministic model than a stochastic one. The OU process is an adequate description of the vast majority of cases (6023). Indeed, the OU process is the best single model for 3462 light curves, with the composite OU process/sinusoid model being the best in 1706 cases. The latter model is the dominant one for brighter/bluer QSOs. Furthermore, a non-negligible fraction of QSO lightcurves show evidence that not only the mean is stochastic but the variance is stochastic, too. Our results confirm earlier work that QSO light curves can be described with a stochastic model, but place this on a firmer footing, and further show that the OU process is preferred over several other stochastic and deterministic models. Of course, there may well exist yet better (deterministic or stochastic) models which have not been considered here.
An efficient separation between dwarfs and giants in surveys of bright stars is important, especially for studies in which distances are estimated through photometric parallax relations. We use the available spectroscopic log g estimates from the sec
ond RAVE data release (DR2) to assign each star a probability for being a dwarf or subgiant/giant based on mixture model fits to the log g distribution in different color bins. We further attempt to use these stars as a labeled training set in order to classify stars which lack log g estimates into dwarfs and giants with a SVM algorithm. We assess the performance of this classification against different choices of the input feature vector. In particular, we use different combinations of reduced proper motions, 2MASS JHK, DENIS IJK and USNO-B B2R2 apparent magnitudes. Our study shows that -- for our color ranges -- the infrared bands alone provide no relevant information to separate dwarfs and giants. Even when optical bands and reduced proper motions are added, the fraction of true giants classified as dwarfs (the contamination) remains above 20%. Using only the dwarfs with available spectroscopic log g and distance estimates (the latter from Breddels et al. 2010), we then repeat the stream search by Klement, Fuchs & Rix (2008, KFR08), which assumed all stars were dwarfs and claimed the discovery of a new stellar stream at V = -160 km/s in a sample of 7015 stars from RAVE DR1. Our re-analysis of the pure DR2 dwarf sample exhibits an overdensity of 5 stars at the phase-space position of the KFR08 stream, with a metallicity distribution that appears inconsistent with that of stars at comparably low rotational velocities. Compared to several smooth Milky Way models, the mean standardized deviation of the KFR08 stream is only marginal at 1.6$pm$0.4... (abbreviated)
We investigate the performance of some common machine learning techniques in identifying BHB stars from photometric data. To train the machine learning algorithms, we use previously published spectroscopic identifications of BHB stars from SDSS data.
We investigate the performance of three different techniques, namely k nearest neighbour classification, kernel density estimation and a support vector machine (SVM). We discuss the performance of the methods in terms of both completeness and contamination. We discuss the prospect of trading off these values, achieving lower contamination at the expense of lower completeness, by adjusting probability thresholds for the classification. We also discuss the role of prior probabilities in the classification performance, and we assess via simulations the reliability of the dataset used for training. Overall it seems that no-prior gives the best completeness, but adopting a prior lowers the contamination. We find that the SVM generally delivers the lowest contamination for a given level of completeness, and so is our method of choice. Finally, we classify a large sample of SDSS DR7 photometry using the SVM trained on the spectroscopic sample. We identify 27,074 probable BHB stars out of a sample of 294,652 stars. We derive photometric parallaxes and demonstrate that our results are reasonable by comparing to known distances for a selection of globular clusters. We attach our classifications, including probabilities, as an electronic table, so that they can be used either directly as a BHB star catalogue, or as priors to a spectroscopic or other classification method. We also provide our final models so that they can be directly applied to new data.
