Effect of training characteristics on object classification: an application using Boosted Decision Trees

Abstract

We present an application of a particular machine-learning method (Boosted Decision Trees, BDTs using AdaBoost) to separate stars and galaxies in photometric images using their catalog characteristics. BDTs are a well established machine learning technique used for classification purposes. They have been widely used specially in the field of particle and astroparticle physics, and we use them here in an optical astronomy application. This algorithm is able to improve from simple thresholding cuts on standard separation variables that may be affected by local effects such as blending, badly calculated background levels or which do not include information in other bands. The improvements are shown using the Sloan Digital Sky Survey Data Release 9, with respect to the type photometric classifier. We obtain an improvement in the impurity of the galaxy sample of a factor 2-4 for this particular dataset, adjusting for the same efficiency of the selection. Another main goal of this study is to verify the effects that different input vectors and training sets have on the classification performance, the results being of wider use to other machine learning techniques.

Accepted for publication in Astronomy & Computing

References in corpus (7)

• Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC
• Boosted Decision Trees as an Alternative to Artificial Neural Networks for Particle Identification
• The Blanco Cosmology Survey: Data Acquisition, Processing, Calibration, Quality Diagnostics and Data Release
• Robust Machine Learning Applied to Astronomical Datasets I: Star-Galaxy Classification of the SDSS DR3 Using Decision Trees
• Studies of Boosted Decision Trees for MiniBooNE Particle Identification
• How to Find More Supernovae with Less Work: Object Classification Techniques for Difference Imaging
• Excess Clustering on Large Scales in the MegaZ DR7 Photometric Redshift Survey