“Fink: Early supernovae Ia classification using active learning”

Context. The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will produce a continuous stream of alerts made of varying sources in the sky. This data flow will be publicly advertised and distributed to scientists via broker systems such as FINK, whose task is to extract scientific information from the stream. Given the complexity and volume of the data to be generated, LSST is a prime target for machine learning (ML) techniques. One of the most challenging stages of this task is the construction of appropriate training samples which enable learning based on a limited number of spectroscopically confirmed objects. Aims. We describe how the FINK broker early supernova Ia (SN Ia) classifier optimizes its ML classifications by employing an active learning (AL) strategy. We demonstrate the feasibility of implementing such strategies in the current Zwicky Transient Facility (ZTF) public alert data stream. Methods. We compared the performance of two AL strategies: uncertainty sampling and random sampling. Our pipeline consists of three stages: feature extraction, classification, and learning strategy. Starting from an initial sample of ten alerts, including five SNe Ia and five non-Ia, we let the algorithm identify which alert should be added to the training sample. The system was allowed to evolve through 300 iterations. Results. Our data set consists of 23 840 alerts from ZTF with a confirmed classification via a crossmatch with the SIMBAD database and the Transient Name Server (TNS), 1600 of which were SNe Ia (1021 unique objects). After the learning cycle was completed, the data configuration consisted of 310 alerts for training and 23 530 for testing. Averaging over 100 realizations, the classifier achieved ~89% purity and ~54% efficiency. From 01 November 2020 to 31 October 2021 FINK applied its early SN Ia module to the ZTF stream and communicated promising SN Ia candidates to the TNS. From the 535 spectroscopically classified FINK candidates, 459 (86%) were proven to be SNe Ia. Conclusions. Our results confirm the effectiveness of AL strategies for guiding the construction of optimal training samples for astronomical classifiers. It demonstrates in real data that the performance of learning algorithms can be highly improved without the need of extra computational resources or overwhelmingly large training samples. This is, to our knowledge, the first application of AL to real alert data.