# Repeated k-fold cross-validation vs. repeated holdout cross-validation: which approach is more reasonable?

I want to split my data 100 times (1/5 as testing, 4/5 as training), and the use the training data to build a model and the testing data to calculate the MSE.

There are two ways we can do this:

1. Do 5-fold cross validation 20 times, i.e., each time samples are split into 5 folds, and each fold will be used as testing dataset.

2. Randomly choose 1/5 of the data as testing set, the other as training set. Do this 100 times.

Which one is more reasonable? Is there a theory of cross-validation that provides a reason to prefer one or the other?

Which method is more reasonable depends on what conclusion you exactly want to draw.

Actually, there is a 3rd possibility which differs from your version 2 by choosing the training data with replacement. This is closely related to out-of-bootstrap validation (differs only by the number of training samples you draw).

Drawing with replacement is sometimes preferred over the cross validation methods as it is closer to reality (drawing a sample in practice does not diminish the chance to draw another sample of the same characteristics again - at least as long as only a very small fraction of the true population is sampled).

I'd prefer such an out-of-bootstrap validation if I want to conclude on the model performance that can be achieved if the given algorithm is trained with $n_{train}$ cases of the given problem. (Though the caveat of Bengio, Y. and Grandvalet, Y.: No Unbiased Estimator of the Variance of K-Fold Cross-Validation Journal of Machine Learning Research, 2004, 5, 1089-1105 does also apply here: you try to extrapolate from one given data set onto other training data sets as well, and within your data set there is no way to measure how representative that data set actually is)

If, on the other hand, you want to estimate (approximately) how good the model you built on the whole data set performs on unknown data (otherwise of the same characteristics of your training data) then I'd prefer approach 1 (iterated/repeated cross validation).

• Its surrogate models are a closer approximation to the model whose performance you actually want to know - so less randomness in the training data is on purpose here.
• The surrogate models of iterated cross validation can be seen as perturbed (by exchanging a small fraction of the training cases) versions of each other. Thus, changes you see for the same test case can directly be attributed to model instability.

Note that whatever scheme you chose for your cross- or out-of-bootstrap validation, you only ever test as much as $n$ cases. The uncertainty caused by a finite number of test cases cannot decrease further, however many bootstrap or set validation (your approach 2) or iterations of cross validation you run.

The part of the variance that does decrease with more iterations/runs is variance caused by model instability.

In practice, we've found only small differences in total error between 200 runs of out-of-bootstrap and 40 iterations of $5$-fold cross validation for our type of data: Beleites et al.: Variance reduction in estimating classification error using sparse datasets, Chemom Intell Lab Syst, 79, 91 - 100 (2005). Note that for our high-dimensional data, resubstition/autoprediction/training error easily becomes 0, so the .632-bootstrap is not an option and there is essentially no difference between out-of-bootstrap and .632+ out-of-bootstrap.

For a study that includes repeated hold out (similar to your approach2), see Kim: Estimating classification error rate: Repeated cross-validation, repeated hold-out and bootstrap Computational Statistics & Data Analysis , 2009, 53, 3735 - 3745.

First you have to understand that the underlying aim of cross validation is to predict the efficiency of a method that is build to classify any future data i.e. in a way compare the efficiency of 2 or more methods (if any). As we do not know any future data, we have to make the method efficient in such a way that it gives maximum efficiency for any random future data. So in a way, we have to implement maximum randomness in our cross validation.

Logically speaking, the second one is more reasonable , as if you take fixed 5 sets as the testing sets, then practically you are cutting down randomness to a great extent as you are fixing which set of data is to be the basis of the model(the training set) and which not. However if you are generating a random process, to select your training set ,then true randomness is implemented. so in a way "truer" application of cross validation will be implemented for your second process.

However everything that has pros also has cons. After all ,keep in mind that cross validation has to be done on the entire data set, which is why you do cross validation 20 times to exhaust the entire data set. 100 times repeating the second process will not(actually with very low probability) exhaust the entire data set. So this is a con of too much randomness.

• Thanks for the answer. I'm with you on the first two paragraphs, but I'm not quite sure I get the last paragraph. What do you mean by 'exhausting the entire data set'? And why the first approach could exhaust while the second couldn't ?
– Lii
Dec 3, 2015 at 19:41
• What do you do in general? Considering the first procedure, divide the data set into 5 parts, repeat the same model procedure 5 times, each time taking a different part as the testing set and the rest 4 parts as the training set. In this way you exhaust the entire data set i.e. you are ensuring each data of the given data set has been tested at least once. And one request, if you like my answer please up-vote . Thanks..:-) Dec 3, 2015 at 20:44
• I got your idea. For now it feels like both are reasonable to some extent. I would appreciate it if you could provide further literatures on this @Qwerty
– Lii
Dec 4, 2015 at 3:19
• @Qwerty: repeating a $5$-fold CV 20 times typically does not exhaust all possible splits into 5 subsets. Dec 4, 2015 at 16:31
• Isn't that(i.e.no change in the process in which data is being collected) an assumption one always has to make while devising a model based on a training set and checking it on a testing set to judge its accuracy? After all, what accuracy do we measure? With how much probability my model will correctly predict(or classify) future data. So the future data must be very similarly collected as the given data. Am I right? Dec 4, 2015 at 19:53