Verify upstream labels

[bfhealy]

On Fritz, the sitewide and scope phenomenological taxonomies have hierarchical connections that are not always maintained in the golden dataset. For example, 93 objects are labeled as periodic but not variable. While that is a small number compared to the ~10⁴ variable sources, for some classes it may be a more substantial fraction. Verifying the upstream labels may improve training results and maintain consistency going forward.

[bfhealy]

The most common missing upstream labels are as follows:

• periodic (3374)
  • Most of the sources missing the periodic label have the long timescale (2936) or wrong period (402) labels.
• irregular (450)
  • Mostly due to flaring (447) sources
• variable (201)
  • 93 periodic sources, various others

There are also 579 non-variable labels missing, mostly from sources labeled as bogus (502). The non-variable label is less important than the ones above, since for 98.8% of sources it is simply 1 - P(variable) and does not train a binary classifier.

The periodic label has the most missing sources, but its classifier is already achieving ~95% precision and recall with the current training set. Going forward, we may want to label flaring sources as irregular variables, since it seems to be an appropriate umbrella term for both flaring and dipping stars.

[park2454]

Hi Brian, I will look into this issue.

[bfhealy]

Thanks Sungmin!