Supernova scores for active anomaly detection

Abstract

Large time-domain sky surveys generate extensive multi-year catalogs of light curves in which scientifically valuable transients, such as supernovae (SNe), are vastly outnumbered by artifacts and routine star variability. While supervised machine learning models can efficiently filter known classes, they struggle with extreme class imbalance and may overlook rare or novel events. Conversely, unsupervised anomaly detection provides broad discovery potential but lacks targeted sensitivity. We present a hybrid strategy that integrates a supervised SN probability score (SN-score) into the PineForest active anomaly detection framework to enhance SN discovery rate in the 23rd data release of the Zwicky Transient Facility. We train a binary classifier using light-curve features of spectroscopically confirmed SNe from the ZTF Bright Transient Survey, achieving a ROC-AUC approximately 0.98. Incorporating the SN-score as an additional feature, together with a small set of labeled priors, significantly accelerates the discovery of SN-like transients across ten extragalactic ZTF fields. This method increases discovery efficiency without compromising the ability to identify diverse astrophysical anomalies. Application of the combined methodology resulted in the discovery of seven previously unreported SN candidates, one AGN candidate, one unusual Galactic variable star SNAD283, as well as two host galaxies exhibiting multiple supernova events. These results demonstrate its value for scalable and expert-guided transient search in current and future surveys, including the Vera C. Rubin Observatory Legacy Survey of Space and Time.