学术文献库

In this work, we assess the combined use of Gaia photometry and astrometry with infrared data from CatWISE in improving the identification of extragalactic sources compared to the classification obtained using Gaia data. We evaluate different input feature configurations and prior functions, with the aim of presenting a classification methodology integrating prior knowledge stemming from realistic class distributions in the universe. In our work, we compare different classifiers, namely Gaussian Mixture Models (GMMs), XGBoost and CatBoost, and classify sources into three classes - star, quasar, and galaxy, with the target quasar and galaxy class labels obtained from SDSS16 and the star label from Gaia EDR3. In our approach, we adjust the posterior probabilities to reflect the intrinsic distribution of extragalactic sources in the universe via a prior function. We introduce two priors, a global prior reflecting the overall rarity of quasars and galaxies, and a mixed prior that incorporates in addition the distribution of the these sources as a function of Galactic latitude and magnitude. Our best classification performances, in terms of completeness and purity of the galaxy and quasar classes, are achieved using the mixed prior for sources at high latitudes and in the magnitude range G = 18.5 to 19.5. We apply our identified best-performing classifier to three application datasets from Gaia DR3, and find that the global prior is more conservative in what it considers to be a quasar or a galaxy compared to the mixed prior. In particular, when applied to the pure quasar and galaxy candidates samples, we attain a purity of 97% for quasars and 99.9% for galaxies using the global prior, and purities of 96% and 99% respectively using the mixed prior. We conclude our work by discussing the importance of applying adjusted priors portraying realistic class distributions in the universe.