o d c oss va dat o , w c
s a so
a ed as t e
otat o
n approach, divides a data set into K subsets and builds up K
Devijver and Kittler, 1982]. Each of K models is trained on the K
(subsets) and tested on the reserved one subset, which is called a
ubset, in turn. The final performance is evaluated on all K test
of the K models. Figure 3.15 shows a 3-fold cross-validation
The K-fold cross-validation is commonly implemented using a
ch will rotate the subset of samples to be selected for training and
The other important feature of the K-fold cross-validation
is that each data point will be tested for exactly one time.
A diagram of 3-fold cross-validation. From left to middle, two folds are selected
a model. From right to middle, one fold is selected for testing.
Jackknife test [Quenouille, 1949; Bishop, 1996], which is
y used for a median- or small-sized data set, is a typical case of
d cross-validation when only one data point is reserved for testing
me. Suppose there are N data points. Rather than constructing K
sing the K-fold cross-validation, N models will be constructed
Jackknife test approach for the generalisation test of a supervised
n important benefit of the Jackknife test is its robustness because
o uncertainty when using the Jackknife test. There is no random
process regarding data division in a Jackknife test process.
mple
classification data [Ajaz and Hussain, 2015] was used for the
ation of how evaluation and generalisation are used. Suppose the