Hi,

No, this is not cross-validation. The method you mentioned called "holdout" method.

However, applying cross-validation helps avoiding overfitting problem.

HTH.

Dr. Samer Sarsam

This does mean that no cross-validation is done.

You can still resolve the overfitting problem using your 75% - 25% data split. Calculate the error (mean square error, misclassification error, whatever is suitable for your problem) for the model you get trained on 75% of the data. Then calculate the error when you use that model to predict the remaining 25%.

If the first error is small and the second is large, then your model doesn't generalize well to unseen examples, and you are overfitting. So you relax the model complexity (usually number of model parameters or depth of tree), and try again. Once the two errors are similar, you are no longer overfitting.

This answer is a simplification that ignores many important aspects (how you do the 75-25 split, model selection, external validation, etc.), but the answer to your question is that cross-validation isn't required to address the problem of overfitting.

The mentioned method is validation but not cross validation. In the validation process there are several methods can be implemented such as holdout, leave one out, cross validation and so on. The method you mentioned is holdout or sometimes called leave-group-out.

Although validation and calibration has been well defined, in the literature there is a misuse of the terms as some time validation used instead of test and calibration used instead of both calibration and validation or sometimes training used instead of calibration and validation.

To expand the mis use of the terms, assuming calibration is training and validation, 75% of the data can be used just for calibrating the model and 25% for validation (i.e. the validation is to prevent over fitting). But this does not leave any data to be used for testing the model after being trained.

If we assume calibration is calibrating and overfitting was prevented with the used algorithm, then the validation here is to test the models. So, 75% is used for the calibration which is calibrating and over fitting prevention (i.e. applied by the algorithm which depends on the used software) and 25% is used for validation which is here used as testing the model.

Therefore, I prefer using three terms (calibration, validation, and testing) and these terms should be clearly defined in any study. Calibration is for calibrating the model, validation is for preventing the over fitting, and testing is to test the model with data never used in the training process and they are only for testing the model with real situation.

There are some terms which used interchangeably such as training which can be used instead of calibration but in my opinion if the term training used it should cover calibration and validation.

So, your question comes from the misuse of the terms in the literature. Most of the algorithms used in the modeling uses the 75% for training (i.e. calibration and validation) and 25% for testing. So, we have to make sure about the algorithm used and how it behaves with the data. Also, we have to be sure that the validation (preventing over fitting) is not testing the model (as this is mistakenly used in the literature) but it is part of the training of the model.

Thank you very much all for all your contribution i really appreciated most.

Regards:@ Sinan Jasim Hadi , Thomas W Kelse and Samer Sarsam