I have a couple of questions regarding the real/bogus classifier. From what I understood from Tatiana Acero-Cuellar’s talk at Rubin Community Workshop 2025, the classification of variable stars requires some work. The current accuracy score for them is 4%. My questions are:
What is the plan to improve accuracy for genuine variable stars?
Will all alerts be released with the classification of real/bogus, or only the ones classified as real?
Are there plans to introduce a better model before the LSST alerts start?
Hi @kkruszynska, thanks for your excellent question. We have a few plans in the works to improve classification performance for variables, which we knew were unlikely to perform well due to the limitations of the training set for ComCam and DP1. We expect to augment the training set with examples from existing catalogs, human labelled examples, and/or injected synthetic sources. The updated model is unlikely to be ready at the time of the first LSST alerts but should follow in the next few months. Our general plan is to allow users to make their own cuts on ML score, but we may restrict very low-probability alerts at first as we and the brokers get up to speed.