Automated physical classification in the SDSS DR10: A catalogue of candidate quasars

We discuss whether modern machine learning methods can be used to characterize the physical nature of the large number of objects sampled by the modern multiband digital surveys. In particular, we applied the MLPQNA (Multi Layer Perceptron with Quasi Newton Algorithm) method to the optical data of the Sloan Digital Sky Survey (SDSS) Data Release 10, investigating whether photometric data alone suffice to disentangle different classes of objects as they are defined in the SDSS spectroscopic classification. We discuss three groups of classification problems: (i) the simultaneous classification of galaxies, quasars and stars; (ii) the separation of stars from quasars; (iii) the separation of galaxies with normal spectral energy distribution from those with peculiar spectra, such as starburst or star-forming galaxies and AGN. While confirming the difficulty of disentangling AGN from normal galaxies on a photometric basis only, MLPQNA proved to be quite effective in the three-class separation. In disentangling quasars from stars and galaxies, our method achieved an overall efficiency of 91.31 per cent and a QSO class purity of similar to 95 per cent. The resulting catalogue of candidate quasars/AGNs consists of similar to 3.6 million objects, of which about half a million are also flagged as robust candidates, and will be made available on CDS VizieR facility.