1. Classification of GAIA data
Coryn A.L. Bailer-Jones
Max-Planck-Institut für Astronomie, Heidelberg
Overview
GAIA classification objectives and available data
Approaches to classification: principles and problems
Example classification using RVS-like data
Some specific issues
Summary

2. GAIA classification objectives
discrete classification of objects
as star, galaxy, quasar, solar system object, supernovae etc.
determination of astrophysical parameters (APs) for stars
Teff, logg, [Fe/H], [/Fe], CNO, A(), Vrot, Vrad, activity
combination with parallax to determine stellar:
luminosity, radius, (mass, age)
identification of unresolved binaries (and parametrization of components where possible)
efficient identification of new types of objects
Goal: catalogue of object classifications and astrophysical parameters

3. GAIA data BBP: 4+ broad band filters all objects
MBP: medium band filters all objects
 object classification; stellar Teff, logg, [Fe/H], A()
RVS: nm spectrum, ~ 0.04 nm/pixel G<17
 stellar Vrad, Vrot, specific element abundances
Astrometry  parallax, kinematics, unresolved binaries
Time domain  ~50 epochs over 5 years (photometric variability)
 Inhomogeneous data
“Redshift”problem: to get RV, need correct SpT template, but
to determine SpT (may) need to know  shift
 use MBP data to give SpT and iterate
Generally: use MBP data to give initial classification of RVS data

4. Classification principles
“Supervised” approach:
use pre-classified data (templates) to infer the desired mapping
apply mapping to any new data to give APs or classes
But, the desired mapping is generally degenerate...

7. Minimum Distance Methods (MDMs)
Search for nearest neighbours (templates) in data space
Assign parameters according to these
Generally interpolate:
either in data space:  = f(d; w)
or in parameter space: D = g(; w)
Need to scale data dimensions
e.g. k-nn, 2 min, cross-correlation
a local classification method
 astrophysical parameter(s)
d1,d2 data
D distance to a template

9. Classification principles
selecting just local neighbours in data space can lead to systematic errors or missed solutions
need to find global (forward) mapping and identify degenerate regions
more complex in higher dimensional spaces (data or parameters)
severity of degeneracy depends upon the density of template grid and noise in the data

10. Artificial Neural Networks (ANNs)
As with MDM,
degeneracy is a problem
Functional mapping: astrophysical parameters = f(data; weights)
Weights determined by training on pre-classified data (templates)
 least squares minimization of total classification error (numerical methods)
 global interpolation of data

11. Classification example with high-res spectra
Database of 611 real stellar spectra from Cenarro et al. (2001)
variation over Teff, logg, [Fe/H]
coverage: nm (same as GAIA RVS)
resolution: nm/pixel (poorer than GAIA?)
SNR: median=70; 90% in range
Randomly split data set into two sets:
train a neural network on one set and test its performance on the other.

12. Distribution over APs in Cenarro et al. data
blue = training data (300)
red = test data (311)

13. Results: Teff and logg

14. Results: [Fe/H]

15. Requirements of the classification scheme
produce both discrete classification and continuous parametrization (e.g. star vs. quasar, APs of stars)
recognition of degeneracies in presence of noise
(i.e. recognise multiple classifications for given data vector)
robustly handle missing and censored data
possible RVS lossy compression (as function of magnitude)
 handle different amounts/formats of data
reliable determination of parametrization uncertainties
accommodate ever-improving stellar models
all this for a very wide range of type of objects ...

16. Classification schemes
Hierarchical
Parallel
P = probability; APs = astrophysical parameters

17. Model training Real spectra and synthetic spectra not identical:
systematic differences (modelling uncertainties, e.g. opacities)
increased cosmic scatter in real spectra (unaccounted-for APs)
1. Can synthetic spectra be used to reliably parametrize GAIA data?
2. Are performances representative of what can be achieved?
3. Do synthetic spectra give the best optimization of phot/spec systems?
2+3 require accurate synthetic spectra (or large set of real spectra)
Can overcome mismatch problem for (1):
use real GAIA data of pre-selected targets to apply corrections to synthetic SEDs
APs of these targets determined from higher resolution spectra from ground-based spectra

18. Summary classification with GAIA data is a challenging problem
methods used so far in (astronomical) classification literature are suboptimal for this purpose
 further development of methods is a high priority
particular problems to overcome are:
- degeneracy (especially with MBP data and compressed RVS data)
- inhomogeneous data
development of classification methods is very dependent on appropriate data (real or synthetic)
- both of targets of interest
- and of “contaminating” objects

19. ICAP: the GAIA classification working group
WG responsible for addressing classification issues for GAIA
14 core members; 17 associate members
GAIA Classification meeting
2-3 December
Heidelberg, Germany
Anyone interested in classification issues
broadly related to GAIA is welcome to attend