Classification and Ranking of Fermi LAT Gamma-ray Sources from the 3FGL Catalog using Machine Learning Techniques

TL;DR

This study employs machine learning techniques, especially Random Forest, to classify gamma-ray sources from the 3FGL catalog as pulsars or AGN with over 96% accuracy, and further distinguishes pulsar subtypes, aiding follow-up and population analysis.

Contribution

The paper introduces the application of machine learning algorithms, notably Random Forest, for high-accuracy classification and ranking of gamma-ray sources and pulsar subtypes in the 3FGL catalog, with validation on known sources.

Findings

Random Forest achieves >96% accuracy in classifying AGN and pulsars.

The models accurately distinguish young and millisecond pulsars with ~90% accuracy.

Predictions assist in identifying unassociated sources and potential X-ray counterparts.

Abstract

We apply a number of statistical and machine learning techniques to classify and rank gamma-ray sources from the Third Fermi Large Area Telescope (LAT) Source Catalog (3FGL), according to their likelihood of falling into the two major classes of gamma-ray emitters: pulsars (PSR) or Active Galactic Nuclei (AGN). Using 1904 3FGL sources that have been identified/associated with AGN (1738) and PSR (166), we train (using 70% of our sample) and test (using 30%) our algorithms and find that the best overall accuracy (>96%) is obtained with the Random Forest (RF) technique, while using a logistic regression (LR) algorithm results in only marginally lower accuracy. We apply the same techniques on a sub-sample of 142 known gamma-ray pulsars to classify them into two major subcategories: young (YNG) and millisecond pulsars (MSP). Once more, the RF algorithm has the best overall accuracy (~90%), while a boosted LR analysis comes a close second. We apply our two best models (RF and LR) to the entire 3FGL catalog, providing predictions on the likely nature of {\it unassociated} sources, including the likely type of pulsar (YNG or MSP). We also use our predictions to shed light on the possible nature of some gamma-ray sources with known associations (e.g. binaries, SNR/PWN). Finally, we provide a list of plausible X-ray counterparts for some pulsar candidates, obtained using Swift, Chandra, and XMM. The results of our study will be of interest for both in-depth follow-up searches (e.g. pulsar) at various wavelengths, as well as for broader population studies.