Research and characterisation of blazar candidates among the Fermi/LAT 3FGL catalogue using multivariate classifications

Context. In the recently published 3FGL catalogue, the Fermi/LAT collaboration reports the detection of gamma-ray emission from 3034 sources obtained after four years of observations. The nature of 1010 of those sources is unknown, whereas 2023 have well-identified counterparts in other wavelengths. Most of the associated sources are labelled as blazars (1717/2023), but the BL Lac or FSRQ nature of 573 of these blazars is still undetermined. Aims. The aim of this study was two-fold. First, to significantly increase the number of blazar candidates from a search among the large number of Fermi/LAT 3FGL unassociated sources (case A). Second, to determine the BL Lac or FSRQ nature of the blazar candidates, including those determined as such in this work and the blazar candidates of uncertain type (BCU) that are already present in the 3FGL catalogue (case B). Methods. For this purpose, multivariate classifiers - boosted decision trees and multilayer perceptron neural networks - were trained using samples of labelled sources with no caution flag from the 3FGL catalogue and carefully chosen discriminant parameters. The decisions of the classifiers were combined in order to obtain a high level of source identification along with well controlled numbers of expected false associations. Specifically for case A, dedicated classifications were generated for high (vertical bar b vertical bar > 10 degrees) and low (vertical bar b vertical bar <= 10 degrees) galactic latitude sources; in addition, the application of classifiers to samples of sources with caution flag was considered separately, and specific performance metrics were estimated. Results. We obtained a sample of 595 blazar candidates (high and low galactic latitude) among the unassociated sources of the 3FGL catalogue. We also obtained a sample of 509 BL Lacs and 295 FSRQs from the blazar candidates cited above and the BCUs of the 3FGL catalogue. The number of expected false associations is given for different samples of candidates. It is, in particular, notably low (similar to 9/425) for the sample of high-latitude blazar candidates from case A.