Model validation performance

Muhammed_Fatih_

Hello together, 

which validation performance (with regard to learning and testing phase) of classification models is quicker? Cross-calidation or the classical split validation (with a 70:30 split)? 

Thank you in advance for your help! 

Best regard, 

Fatih

varunm1

Split validation is quicker, it builds model only once and then tests on the dataset. Incase of cross-validation, the model is built k+1 times (K is the number of folds).

I didn't encounter any special case where cross-validation performed faster than split. I don't think it happens if all other settings are the same (Feature selection, hyperparameters, etc).

Maybe if you use a processor with multiple cores and each cross-validation process is run parallelly, then there might be a chance based on the fold sizes. But in general, the above is fine.

Hope this helps.

MartinLiebig

Hi,

Cross-validation is the more accurate estimator of the true model performance.

Best,

Martin

varunm1

Split validation is quicker, it builds model only once and then tests on the dataset. Incase of cross-validation, the model is built k+1 times (K is the number of folds).

I didn't encounter any special case where cross-validation performed faster than split. I don't think it happens if all other settings are the same (Feature selection, hyperparameters, etc).

Maybe if you use a processor with multiple cores and each cross-validation process is run parallelly, then there might be a chance based on the fold sizes. But in general, the above is fine.

Hope this helps.

Muhammed_Fatih_

Hi varunm, 

thank you four your answer! An additional question - is there a possibility to say that one of the two validation processes (Split Validation vs. Cross-Validation) performs better in general with regard to learning and testing? 

Best regards! 

MartinLiebig

Hi,

Cross-validation is the more accurate estimator of the true model performance.

Best,

Martin