Insights into Quasar UV Spectra Using Unsupervised Clustering Analysis

Authors

A Tammour, Department of Physics and Astronomy, University of Western Ontario, 1151 Richmond St, London, ON N6A 3K7, Canada
S.C. Gallagher, Department of Physics and Astronomy, University of Western Ontario, 1151 Richmond St, London, ON N6A 3K7, Canada; Centre for Planetary and Space Exploration, University of Western Ontario,1151 Richmond St, London, ON N6A 3K7, Canada
M Daley, Department of Computer Science, University of Western Ontario,1151 Richmond St, London, ON N6A 3K7, Canada
G.T. Richards, Department of Physics, Drexel University, Philadelphia, PA 19104, USA

Document Type

Article

Publication Date

6-2016

Journal

Monthly Notices of the Royal Astronomical Society

Volume

459

Issue

2

First Page

1659

Last Page

1681

URL with Digital Object Identifier

10.1093/mnras/stw586

Abstract

Machine learning techniques can provide powerful tools to detect patterns in multidimensional parameter space. We use K-means - a simple yet powerful unsupervised clustering algorithm which picks out structure in unlabelled data - to study a sample of quasar UV spectra from the Quasar Catalog of the 10th Data Release of the Sloan Digital Sky Survey (SDSS-DR10) of Paris et al. Detecting patterns in large data sets helps us gain insights into the physical conditions and processes giving rise to the observed properties of quasars. We use K-means to find clusters in the parameter space of the equivalent width (EW), the blue- and red-half-width at half-maximum (HWHM) of the Mg II 2800 Å line, the C IV 1549 Å line, and the C III] 1908 Å blend in samples of broad absorption line (BAL) and non-BAL quasars at redshift 1.6-2.1. Using this method, we successfully recover correlations well-known in the UV regime such as the anti-correlation between the EW and blueshift of the C IV emission line and the shape of the ionizing spectra energy distribution (SED) probed by the strength of He II and the Si III]/C III] ratio. We find this to be particularly evident when the properties of C III] are used to find the clusters, while those of Mg II proved to be less strongly correlated with the properties of the other lines in the spectra such as the width of C IV or the Si III]/C III] ratio. We conclude that unsupervised clustering methods (such as K-means) are powerful methods for finding `natural' binning boundaries in multidimensional data sets and discuss caveats and future work.