In classification approaches,how to select the value of K for K-fold cross validation?..Found that arbitrary value of K is 10 in almost all literature. In some cases k=5.. Why to choose 10 & Why not >10?
Hello,
if you assume that 50% of data is enough for training a classifier, often it is performed 5 × 2 𝑐𝑣 𝐹 cross-validation method (meaning five times 2-
fold cross-validation). The original of the method is cited below. It was pointed out in the paper that after five folds, the sets
overlap and share many statistics implying that new folds do
not add new information.
T. G. Dietterich, “Approximate Statistical Tests for Comparing Supervised
Classification Learning Algorithms,” Neural Computation, vol. 10,
pp. 1895–1923, 1998.
This is something that is mostly chosen empiricaly, taking into account facts like: the number of instances, number of variables and their types, the number of your outcome classes and their ratio, in other words how balanced or imbalanced your dataset is regarding the classes. In case of an imbalanced dataset you should also consider using a Stratified Cross-Validation. You can find more information about Stratification and its effects in our paper. Now regarding the "why not >10". Well we mostly apply Cross-Validation when we don't have an abundance of data so most probably that would mean too few data in each fold and that in turn most of the times translates to increased variance.
Conference Paper Performance-Estimation Properties of Cross-Validation- Based...
- Badges
- Science topic
- Similar topics
- Social Psychology
- Values