Page 248 - Special Topic Session (STS) - Volume 1

STS426 Didier Fraix-Burnet

Unsupervised classification of galaxy spectra and
interpretability
Didier Fraix-Burnet
Univ. Grenoble Alpes / CNRS / IPAG, Grenoble, France

Abstract
Dealing with large amount of data is a new problematic task in astrophysics.
One may distinguish the management of these data (astroinformatics) and
their scientific use (astrostatistics) even if the border is rather fuzzy.
Dimensionality reduction in both the number of observations and the number
of variables (observables) is necessary for an easier physical understanding.
This is the purpose of classification which has been traditionally eye-based and
essentially still is, but this becomes not possible anymore. In this talk, I present
an unsupervised classification of 700 000 spectra of galaxies of 1500
wavelengths each, with a model-based subspace clustering algorithm (Fisher-
EM). I also show some preliminary results on the interpretation of the classes
using data bases of modelled spectra.

Keywords
Unsupervised classification; machine learning; spectra; astrophysics; galaxies

1. Introduction
Astrophysics has now entered the era of Big Data and the new telescopes
and instruments that will come into operation in the next few years (EUCLID,
VLT/MOONS, LSST, SKA...) face technological challenges for the management
and the analysis of the data. Spectra are particularly spectacular since they
contain several thousands of wavelengths making matrices of about a million
observations described by thousands of parameters.
These spectra contain all the astrophysical information that an astronomer
can dream of, apart from the morphological structure: the composition of the
stellar populations, the history of the stellar formation events, the content in
gas and its physical conditions, the presence of hot regions such star forming
regions, hot nebulae, active galactic nuclei hosting black holes, and the global
kinematics of the galaxy. Basically the spectrum of a galaxy is made of a
continuum due to the thermal emission from the stars, plus some absorption
features due to the cold gas, and emission lines due to hot gas. An atlas of
typical galaxy spectra is provided in Kennicutt (1992) or Dobos et al. (2012).
Classification in astrophysics traditionally uses an eye-based approach
which also serves as the bases for supervised learning studies. In contrast,
unsupervised learning is not common (Fraix-Burnet et al. 2015). In the case of

237 | I S I W S C 2 0 1 9