Taxonomy overclassification and underclassification
Taxonomy benchmark home
Validating taxonomy classifiers
Training taxonomy classifiers
Splitting a taxonomy reference set
Defining "accuracy" of a taxonomy classifier
Taxonomy classification errors
Lowest Common Rank (LCR)
The lowest common rank is defined to be the lowest
level (genus, family etc.) that is present in the reference set (training
set). For example, if the
genus is not present but the family is present, then the lowest common level is
family. Predicting the LCR is the most difficult challenge for a taxonomy
classification algorithm. For example, if the identity of the top hit is 88%,
should be LCR be class, family, genus or what?
See also overclassification benchmark results.
If a classifier predicts a level lower than the LCR, then this is an
overclassification error. For example, the lowest common level is family but a
genus is predicted.
If a classifier predicts higher levels but does not predict the LCR, then this is an
underclassification error. For example, the lowest common level is family but
the lowest predicted level is order. All false negatives are underclassification
errors so there is really no need for a new term.