Soft Labels

  • What are soft labels?

By default labels are crisp: every object can be assigned to just a single class. Soft labels give the possibility to assign class memberships to all classes. Soft labels don’t necessarily sum to one for a single object. Routines that expect that may use normm for such a normalization. Not all routines in PRTools can handle soft labels.

  • Conversion of a crisp labeled dataset into a soft labeled one doesn’t seem to work.

Labels can be inspected by getlabels. For a crisp labeled dataset just a set of m labels is shown. A soft labeled dataset will display a m x c set of the c class memberships for the m objects. As PRTools tries to be consistent, the existing soft labeled implementations are preferably programmed such that they are consistent with crisp labels in case the soft labels have a single class memberships of 1.