Tags for false positive classes

[AnderssonAstro]

I know this is not the primary intention of the pipeline, but having tags for common false positive classes(e.g. sidelobes, bad subtractions, satellites(?)) would be a useful endeavour. For example, quantifying what fraction of false positives are caused by what issue is good to know for future developments and tracking the stats ("95% of our false positives happen around bright sources" and similar). Similarly for machine learning endeavours, having labelled classes beyond just 'false positive' would allow for more multivariate classification schemes.

[Elle-WANG]

Good suggestion! The first version of the tags included some false-positive classes, but we decided to remove them because they can be confusing for new users. We also can’t fully trust the labels rated by them - distinguishing between true and false detections is relatively straightforward, but classifying different types of false positives requires more expert skills.

I agree this feature would be useful for machine learning, but in practice, experts might still need to review the data to ensure a reliable training set.

[AnderssonAstro]

Totally agree, classifying the reals as real is more important by a long way.

One could split it into a two stage classification (i.e. first Real or Bogus and then assigning an appropriate astrophysical/artefactual type) as opposed to the current one-step setup. One could then try a 'simpler' model for classifying into real/bogus, which would have better purity (as you say it's a more straightforward task) and then have a stretch goal of classifying each kind of detection and having to deal with the increase in label contamination.