10.3.5 Quality assessment and validation
Supervised classification required detailed verification and validation with the literature at all stages, from the training of classifiers to the selection of results.
As mentioned in Section 10.3.4, the training set was selected by verification of a comprehensive cross-match with the literature, performed at the most detailed class level, in order to ensure the representation of each (sub)class. Verification included visual inspections of aggregated one- and two-dimensional diagrams (such as the observational Hertzsprung–Russell diagram, combinations of time-series features and other photometric and astrometric parameters), in addition to per-source time series plots, depending on the variability type. The verification and selection of results were in most cases guided by their adherence to the known objects of those classes, similar to the training set construction, but with less severe constraints.
The validation of the classification results employed sources from the cross-match of variable objects in the literature (Gavras et al. 2023) and from the database of the Centre de Donneés astronomiques de Strasbourg (http://cdsarc.u-strasbg.fr/). This assessment was performed for each class and presented in Rimoldini et al. (2023).