Abstract
In this paper we present an experimental study of the performance of a simple machine learning algorithm applied to the prediction of the stellar atmospheric parameters T[eff], log g and [Fe/H] using as input three different sets of spectral features. We compare the performance of the distance-weighted 3-nearest-neighbor algorithm using as input spectra, a set of spectral indices taken from the same wavelength region, and absorption lines obtained by removing from the spectra the contribution of the continuum, which is computed by means of a linear time convex hull algorithm. Our experiments show that the predictions obtained using spectral indices and spectral lines have very similar accuracy levels, and that both are superior to those obtained using spectra.
