Updates for January 10

[bfhealy]

General announcements

SkyPortal upgrades (to be detailed in video this week):

• Sources can be filtered by whether or not they were labelled by the current user (Filter sources labelled by current user skyportal/skyportal#3812)
• Votes for/against classifications are now hidden by default to avoid biasing labellers (No classifications are also non-classifications skyportal/skyportal#3823)
  • Filtering by non-classification received bugfixes and now has the option to return sources that are unlabelled

Pull requests

• Consistent/custom imputation strategy #207
• Facilitate subsequent training for existing models #208

Going forward

All recent SkyPortal feedback from SCoPe scanners has now been incorporated into the app. We will create a SkyPortal Astronomy video detailing the upgrades this week.

Work continues on updating the SCoPe feature imputation strategy. We will move from an inconsistent zero/mean imputation strategy to one that combines zero/median/KNN regression on a feature-by-feature basis. After completing this update, we will be in good position to re-run training and achieve more reliable inference results. Increasing the balance hyperparameter (controlling the ratio of negative/positive examples) should allow the models to better learn the specific details of each class in comparison to the others.