Gradient boosting decision trees classification of blazars of uncertain type in the fourth Fermi-LAT catalog

Abstract

The deepest all-sky survey available in the $γ$-ray band - the last release of the Fermi-LAT catalogue (4FGL-DR3) based on the data accumulated in 12 years, contains more than 6600 sources. The largest population among the sources is blazar subclass - 3743, $60.1\%$ of which are classified as BL Lacertae objects (BL Lacs) or Flat Spectrum Radio Quasars (FSRQs), while the rest are listed as blazar candidates of uncertain type (BCU) as their firm optical classification is lacking. The goal of this study is to classify BCUs using different machine learning algorithms which are trained on the spectral and temporal properties of already classified BL Lacs and FSRQs. Artificial Neural Networks, \textit{XGBoost} and LightGBM algorithms are employed to construct predictive models for BCU classification. Using 18 input parameters of 2219 BL Lacs and FSRQs, we train (80\% of the sample) and test (20\%) these algorithms and find that LightGBM model, state-of-the-art classification algorithm based on gradient boosting decision trees, provides the highest performance. Based on our best model, we classify 825 BCUs as BL Lac candidates and 405 as FSRQ candidates, however, 190 remain without a clear prediction but the percentage of BCUs in 4FGL is reduced to 5.1\%. The $γ$-ray photon index, synchrotron peak frequency, and high energy peak frequency of a large sample are used to investigate the relationship between FSRQs and BL Lacs (LBLs, IBLs, and HBLs).