Machine learning methods for constructing probabilistic Fermi-LAT catalogs
Classification of sources is one of the most important tasks in astronomy. Sources detected in one wavelength band, for example using gamma rays, may have several possible associations in other wavebands, or there may be no plausible association candidates. In this work we aim to determine the probabilistic classification of unassociated sources in the third Fermi Large Area Telescope (LAT) point source catalog (3FGL) and the fourth Fermi LAT data release 2 point source catalog (4FGL-DR2) using two classes - pulsars and active galactic nuclei (AGNs) - or three classes - pulsars, AGNs, and "OTHER" sources. We use several machine learning (ML) methods to determine a probabilistic classification of Fermi-LAT sources. We evaluate the dependence of results on the meta-parameters of the ML methods, such as the maximal depth of the trees in tree-based classification methods and the number of neurons in neural networks. We determine a probabilistic classification of both associated and unassociated sources in the 3FGL and 4FGL-DR2 catalogs. We cross-check the accuracy by comparing the predicted classes of unassociated sources in 3FGL with their associations in 4FGL-DR2 for cases where such associations exist. We find that in the two-class case it is important to correct for the presence of OTHER sources among the unassociated ones in order to realistically estimate the number of pulsars and AGNs. We find that the three-class classification, despite different types of sources in the OTHER class, has a similar performance as the two-class classification in terms of reliability diagrams and, at the same time, it does not require adjustment due to presence of the OTHER sources among the unassociated sources. We show an example of the use of the probabilistic catalogs for population studies, which include associated and unassociated sources.