Page 250 - Special Topic Session (STS) - Volume 1
P. 250
STS426 Didier Fraix-Burnet
3. Results
We here present the result on one set of 100 000 spectra (Fig. 1) as well as
the one for 300 000 spectra (Fig. 2) for comparison. In all cases, the optimal
number of groups according to the integrated completed likelihood (ICL)
criterion is found to be 20.
For the first set, four of the groups have only one or two spectra which are
very odd, while for the second set three groups are in this case. These odd
spectra are either too noisy, or caused by an technical problem at the time of
observation, or coming from another kind of source (star?) and misclassified
in the automatic procedure. They are not considered in this paper and not
shown in the figures, but the fact that they were isolated among the large data
set is in itself a nice outcome since real outliers could thus be found.
Another noticeable result is that both classifications agree very well, the
same kinds of spectra can be easily identified. Indeed, this is true for the results
on all the sets of 100 000 spectra, showing the robustness of the clustering
algorithm. A more quantitative comparison is in progress.
The dispersion of spectra within groups is always low, there is little overlap
implying that the groups are highly specific. This is remarkable since both the
number of observations and the number of variables are very large. In
addition, galaxies are characterized by quantitative properties that vary
continuousy over large ranges of values, so that the data space made by their
spectra could be expected to be more or less homogeneously populated,
without clearcut structures. Our result tends to show that this is not the case
however.
For astronomers, the most remarkable outcome is that the groups map
remarkably well the diversity of galaxy spectra that has been observed or
modelled. This is certainly the first time in astrophysics that without any
physical a priori, a classification is obtained and provides immediate and easily
identifiable physical properties.
239 | I S I W S C 2 0 1 9
