2026ApJ..1000..181T

Fast radio bursts (FRBs) are millisecond-duration radio transients whose physical origin and population structure remain unresolved. A commonly invoked observational distinction-whether a source repeats-has been used to constrain progenitor models, yet this classification is inherently incomplete because the absence of detected repetition does not imply a truly nonrepeating source. Here we cast FRB classification as a positive-unlabeled learning problem and apply a semisupervised machine learning framework to the CHIME/FRB catalog, without presuming that unlabeled sources are genuine nonrepeaters. The resulting classifier successfully recovers most confirmed repeaters and identifies a substantial subset of apparently nonrepeating sources with properties consistent with repetition. Feature-based analyses further reveal that the spectral quality factor and the rest-frame temporal width dominate the population separation. These findings demonstrate that explicitly accounting for label uncertainty is crucial for robust FRB population studies and provides a physically interpretable route toward constraining FRB emission mechanisms.