Imbalanced learning for RR Lyrae stars

TL;DR

This paper presents a machine learning approach combined with convex hull algorithms to effectively identify RR Lyrae stars from large astronomical datasets, demonstrating improved accuracy by leveraging multi-band photometric data and addressing class imbalance.

Contribution

It introduces a novel combination of convex hull and machine learning techniques for RR Lyrae star classification, highlighting the importance of multi-band data and cost-sensitive algorithms.

Findings

Convex hull method achieves up to 16.1% efficiency with 53% completeness.

GALEX ultraviolet data improves RR Lyrae star identification.

Fast Boxes algorithm performs best on imbalanced astronomical data.

Abstract

We apply machine learning and Convex-Hull algorithms to separate RR Lyrae stars from other stars, like main sequence stars, white dwarf stars, carbon stars, CVs and carbon-lines stars, based on the Sloan Digital Sky Survey (SDSS) and Galaxy Evolution Explorer (GALEX). In the low-dimensional space, the Convex-Hull algorithm is applied to select RR Lyrae stars. Given different input patterns of (u-g, g-r), (g-r, r-i), (r-i, i-z), (u-g, g-r, r-i), (g-r, r-i, i-z), (u-g, g-r, i-z) and (u-g, r-i, i-z), different convex hulls can be built for RR Lyrae stars. Comparing the performance of different input patterns, u-g, g-r, i-z is the best input pattern. For this input pattern, the efficiency (the fraction of true RR Lyrae stars in the predicted RR Lyrae sample) is 4.2% with a completeness (the fraction of recovered RR Lyrae stars in the whole RR Lyrae sample) of 100%, increases to 9.9% with 97% completeness and to 16.1% with 53% completeness by removing some outliers. In the high-dimensional space, machine learning algorithms are used with input patterns (u-g, g-r, r-i, i-z), (u-g, g-r, r-i, i-z, r), (NUV-u, u-g, g-r, r-i, i-z) and (NUV-u, u-g, g-r, r-i, i-z, r). RR Lyrae stars, which belong to the class of interest in our paper, are rare compared to other stars. For the highly imbalanced data, cost-sensitive Support Vector Machine (SVM), cost-sensitive Random Forest and Fast Boxes are used. The results show that information from GALEX is helpful for identifying RR Lyrae stars and Fast Boxes are the best performers on the skewed data in our case.