Classifications of Fermi-LAT unassociated sources in multiple machine learning methods

Abstract

ABSTRACT The classifications of Fermi-LAT unassociated sources are studied using multiple machine learning (ML) methods. The update data from 4FGL-DR3 are divided into high Galactic latitude (HGL, Galactic latitude |b| > 10°) and low Galactic latitude (LGL, |b| ≤ 10°) regions. In the HGL region, a voting ensemble of four binary ML classifiers achieves a 91 per cent balanced accuracy. In the LGL region, an additional Bayesian–Gaussian (BG) model with three parameters is introduced to eliminate abnormal soft spectrum active galactic nuclei (AGNs) from the training set and ML-identified AGN candidates, a voting ensemble of four ternary ML algorithms reach an 81 per cent balanced accuracy. And then, a catalogue of Fermi-LAT all-sky unassociated sources is constructed. Our classification results show that (i) there are 1037 AGN candidates and 88 pulsar candidates with a balanced accuracy of 0.918 ± 0.029 in HGL region, which are consistent with those given in previous all-sky ML approaches; and (ii) there are 290 AGN-like candidates, 135 pulsar-like candidates, and 742 other-like candidates with a balanced accuracy of 0.815 ± 0.027 in the LGL region, which are different from those in previous all-sky ML approaches. Additionally, different training sets and class weights were tested for their impact on classifier accuracy and predicted results. The findings suggest that while different training approaches can yield similar model accuracy, the predicted numbers across different categories can vary significantly. Thus, reliable evaluation of the predicted results is deemed crucial in the ML approach for Fermi-LAT unassociated sources.