Prior-Sensitive Ranking of Fast Radio Burst Repeater Candidates for Targeted Follow-Up

Identifying fast radio burst (FRB) sources likely to repeat can guide targeted observations and advance our understanding of their populations. We investigate repeater-candidate selection using 18 observational parameters from a Blinkverse-based sample of 3990 sources, including 104 confirmed repeaters. Eight model configurations are compared across four positive-training sizes and 30 random seeds. Positive–unlabelled (PU) Bagging-RF achieves competitive performance, recovering an average of 92.27% of 44 held-out repeaters within the highest-ranked 30% of test sources when trained on 60 confirmed repeaters. We further combine this model with cross-fitted calibration to estimate conditional repeater probabilities for 3886 unlabelled sources. Analysis across assumed repeater fractions quantifies prior sensitivity, while repeated fits characterize ranking stability. The resulting catalogue integrates these estimates with source-specific observational information to support follow-up target selection.