Positive and unlabelled machine learning reveals new fast radio burst repeater candidates

Abstract

ABSTRACT Fast radio bursts (FRBs) are astronomical radio transients of unknown origin. A minority of FRBs have been observed to originate from repeating sources, and it is unknown which apparent one-off bursts are hidden repeaters. Recent studies increasingly suggest that there are intrinsic physical differences between repeating and non-repeating FRBs. Previous research has used machine learning classification techniques to identify apparent non-repeaters with repeater characteristics, whose sky positions would be ideal targets for future observation campaigns. However, these methods have not sufficiently accounted for the positive and unlabelled (PU) nature of the data, wherein true labels are only available for repeaters. Modified techniques that do not inadvertently learn properties of hidden repeaters as characteristic of non-repeaters are likely to identify additional repeater candidates with greater accuracy. We present in this paper the first known attempt at applying PU-specific machine learning techniques to study FRBs. We train an ensemble of five PU-specific classifiers on the available data and use them to identify 66 repeater candidates in burst data from the CHIME/FRB collaboration, 18 of which were not identified with the use of machine learning classifiers in past research. Our results additionally support repeaters and non-repeaters having intrinsically different physical properties, particularly spectral index, frequency width, and burst width. This work additionally opens new possibilities to study repeating and non-repeating FRBs using the framework of PU learning.