The First Use of Positive and Unlabeled Machine Learning to Identify Fast Radio Burst Repeater Candidates

Fast radio bursts (FRBs) are astronomical radio transients of unknown origin. Some have been observed to repeat, providing valuable insights into their physical origin. It is hypothesized that repeating and non-repeating FRBs physically differ, thus motivating the use of machine learning to identify apparent one-off FRBs with repeater characteristics. Past literature does not sufficiently adapt conventional learning methods to the positive and unlabeled (PU) nature of the data, where true labels are only available for repeaters. We present the first known attempt at applying PU-specific techniques to study FRBs. We apply an ensemble of PU learning classifiers to a catalog of 536 FRBs and identify two previously undiscovered repeater candidates on which to focus future follow-up observations. Our results support repeaters having intrinsic physical differences from non-repeaters and open new research possibilities to study FRBs using the framework of PU learning.