The 35th annual conference on Astronomical Data Analysis Software & Systems 2025 in Görlitz, Germany

We present a performance analysis of the quasar and galaxy classifications provided by the Discrete Source Classifier (DSC) in the Gaia Data Release 3 (GDR3) and propose a new approach to combining the results of individual classifiers to create higher-purity quasar and galaxy catalogues.

In GDR3, the DSC probabilistically classifies sources using a Bayesian framework so that a source is identified as a quasar, a galaxy, a star, a white dwarf or a binary.
The DSC Combmod classifier in GDR3 achieved a high completeness for quasars and galaxies but a low level of purity due to contamination from the far larger star class. To showcase the variations in performance with the sources magnitudes and sky positions, we compute two-dimensional representations of the completeness and the purity as function of Galactic latitude and source brightness. We re-evaluate the GDR3 DSC classifications on a cleaner validation set, excluding the Magellanic Clouds, with no change to the published DSC probabilities.
Here, the GDR3 DSC Combmod achieves average 2-d completenesses of over 92% and average 2-d purities of 55% and 89% for the quasar and galaxy classes, respectively.
A new parametric combination of the baseline classifiers (Allosmod and Specmod), named Combmod-α, achieves an increase in purity by 24 and 5 percentage points but a decrease in completeness by 10 and 2 percentage points, for the quasar and galaxy classes, respectively. At faint magnitudes G≥20, the purity of the quasar class increases by 42 percentage points while the completeness decreases by only 15 percentage points.

Our results demonstrate a significant improvement in purity for a modest loss of completeness for the extragalactic classes, particularly for quasars with faint magnitudes, based solely on a new combination of the published GDR3 DSC probabilities.