A machine learned classifier for RR Lyrae in the VVV survey

¹ Departmento de Estadística, Facultad de Matemáticas, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, 7820436 Macul, Santiago, Chile
e-mail: susana@mat.puc.cl
² Millennium Institute of Astrophysics, 1515 Santiago, Chile
³ Instituto de Astrofísica, Facultad de Física, Pontificia Universidad Católica de Chile, Av. Vicuña Mackenna 4860, 7820436 Macul, Santiago, Chile
⁴ Gemini Observatory, Chile
⁵ Unidad de Astronomía, Facultad Cs. Básicas, Universidad de Antofagasta, Avda. U. de Antofagasta 02800, Antofagasta, Chile
⁶ Departamento de Física, Universidade Federal de Santa Catarina, Trindade 88040-900, Florianópolis, SC, Brazil
⁷ Departamento de Ciencias Físicas, Universidad Andres Bello, República 220, Santiago, Chile
⁸ Vatican Observatory, 00120 Vatican City State, Italy

Received: 12 April 2016
Accepted: 25 August 2016

Abstract

Variable stars of RR Lyrae type are a prime tool with which to obtain distances to old stellar populations in the Milky Way. One of the main aims of the Vista Variables in the Via Lactea (VVV) near-infrared survey is to use them to map the structure of the Galactic Bulge. Owing to the large number of expected sources, this requires an automated mechanism for selecting RR Lyrae, and particularly those of the more easily recognized type ab (i.e., fundamental-mode pulsators), from the 10⁶−10⁷ variables expected in the VVV survey area. In this work we describe a supervised machine-learned classifier constructed for assigning a score to a K_s-band VVV light curve that indicates its likelihood of being ab-type RR Lyrae. We describe the key steps in the construction of the classifier, which were the choice of features, training set, selection of aperture, and family of classifiers. We find that the AdaBoost family of classifiers give consistently the best performance for our problem, and obtain a classifier based on the AdaBoost algorithm that achieves a harmonic mean between false positives and false negatives of ≈7% for typical VVV light-curve sets. This performance is estimated using cross-validation and through the comparison to two independent datasets that were classified by human experts.