About 10-fold cross validation train/test split

Question

So, I want to do 10 fold CV. After I googled it, all of the websites I've found told me that to do the split, take 1 fold as test and the rest as train. But my professor told me another way. She told me to take 1 fold as train and 1 fold as test (and repeat until 10). So in 10 fold, the way she meant is:

1. Model1: Fold1 (Train) + Fold2 (Test)
2. Model2: Fold1 (Train) + Fold3 (Test)

And so on until

9. Model9: Fold1 (Train) + Fold10 (Test)

After that continue to Fold2 as train

10. Model10: Fold2 (Train) + Fold1 (Test)
11. .............. Fold2 (Train) + Fold10 (Test)

It continues until it reaches: Fold10 (Train) + Fold9 (Test). So there is 10x9 iteration until it finished.

I told her the former way (take 1 as test and the rest as train) but she said that the latter way (her way) is more valid/stronger and made the model robust.

So I'm wondering, is her way of doing the CV correct? (Excuse my english, btw)

Answer 1

I think what you describe as $k$-fold cross-validation is fine.

I would urge you to use freely available and established references in your work instead of websites; websites might be excellent at times but it can be hard to convince people of their quality and/or detect "mistakes" when starting in ML. For example, on the matter of cross-validation:

• Hastie et al. (2009) Elements of Statistical Learning, Sect. 7.10 Cross-Validation,
• Shalev-Shwartz & Ben-David (2014) Understanding Machine Learning: From Theory to Algorithms, Sect. 11.2.4 $k$-Fold Cross Validation and
• Bishop (2006) Pattern Recognition and Machine Learning, Sect. 1.3. Model Selection

can all serve as authoritative, well-established and widely used references that will withstand academic scrutiny. For that matter, probably most of the areas covered in an undergraduate ML course will be included in one of these books.

On a purely interpersonal level: Your lecturer might have a particular application in mind. Politely ask him/her to point you to some relevant material.