Page 253 - Special Topic Session (STS) - Volume 1

STS426 Didier Fraix-Burnet
classification is statistically robust through analyses of several subsets of 100
000 observations. This classification appear quite relevant through an eye-
based examination of the median spectra. We have given some hints on the
physical meaning of the spectra and the relevance of the groups and their
specificities. However, a more detailed analysis is required to estimate the
astrophysical properties of the galaxies of the different groups. This work is in
progress.
We want to insist on the fact that the number of groups is chosen thanks
to an objective likelyhood indicator. In addition, this number of groups does
not vary with the different subsets of 100 000 distinct spectra, nor with the 300
000 subset. This probably shows that the true number of typical spectra that
can be distinguished in the data is around 16 or 17 is we exclude the few
outliers that came out in all analyses. This number of typical spectra is given
by statistics, and may or may not correspond to the true number of typical
galaxies. To check this point, careful fits of each spectra will have to be
performed (Moultaka et al. 2004 ; Noll et al. 2009).
Even if it is possible to fit millions of spectra with models and derive
physical properties of galaxies (Comparat et al. 2017), classification remains
invaluable since it allows to perform more precise fits on much fewer typical
spectra. In any case, classification would still be required to simplify and
understand the properties of millions of galaxies and we find highly preferable
to classify the data themselves rather than the derived values prone to their
own uncertainties, errors, degeneracies and model inadequacies.
The present results open the possibility of an automatic and objective
classification procedure for the big databases already available and yet to
come. An extension of the algorithm, called sparse-FEM, allows to identify the
most discriminant parameters (here wavelengths) that explain the
classification. In other words, it is then possible to perform a supervised
classification of new spectra by using these parameters which are much less
(say 100), alllowing for a nearly real time classification.
Finally, an important outcome of our work is the possibility to easily detect
outliers or interesting and rare objects (Baron & Poznanski 2017). The real time
supervised classification could also be used to detect quickly any modification
in the spectrum of a known galaxy. This would be quite useful for the transient
alerts sent to other telescopes.

References
1. Baron, D. & Poznanski, D. (2017). The weirdest SDSS galaxies: results
from an outlier detection algorithm. Monthly Notices of the Royal
Astronomical Society, 465, 4530-4555.

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