k-fold Cross Validation¶
Definition¶
The dataset is partitioned into k disjoint sets called folds. The training set are the k-1 folds, while the test is the last fold.
We use all possible combinations, ie each fold will be used as a test set once.
The advantage of k-fold CV is that the resulting k performance estimates are largely independent of each other, only dependent on the fact theres a slight dependency introduced between training sets.
There is also a little bit of bias introduced since we're not testing on the entire training set, rather we're testing on \(n- (\frac{n}{k})\) of the training set.
The disadvantage of k-fold CV is that \(k\) dictates the number of train-test splits we create and the size of the train and test data.
A common choice for k-fold CV is to choose \(k=10\)