nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Accessing out of fold metrics for K-Fold CV

CentroidError56

Is there any way to access the out of fold metrics when using K-Fold CV with methods like Neural and SVM? As far as I know JMP only reports metrics from the best performing fold, which isn't very useful at all.

Accepted answers

Victor_G

Hi @CentroidError56,

You're right, this option will only work with default hyperparameter settings.
Here is how I would do with specific hyperparameter configurations to be tested on the same folds:

Create a K-folds validation column using the Make Validation Column platform. Specify the stratification columns (the X's) and the target (Y). You can then define the number of folds and set a random seed.
Launch the modeling platform of your choice with this K-folds column in the validation role.
In this example for illustration, I used the JMP Diabetes dataset and the SVM platform. I realized a tuning design to find 3 appropriate hyperparameters settings (models 2, 3 and 20) :
I then press CTRL + click on the option Publish prediction formula (available in the red triangle next to any model). A new Formula Depot window is created with the formula of the three models seen earlier :
In the red triangle next to Formula Depot, I click on the option Model Comparison and select my three models :

JMP notices than my 5-folds validation column is used for all models, and ask if it can use it as a grouping column, so I click yes to have individual models results for each fold.
I have now access to the individual folds measure of performances for each of the three models :

This method should work no matter the modeling platform used, as it is only using the models prediction formula, so you can combine different models types.

Hope this answer will solve your problem,

All comments

Victor_G

Hi @CentroidError56,

Yes, please use the platform Model Screening with the models you want to test checked to get access to validation folds results in the report.

Training, Validation, and Test Measures of Fit

From the Model Screening platform options, you can save a KFolds Results table.

Hope this solution will help you,

CentroidError56

Hi Victor,

Thanks for the response. As far as I can see, this only works with default settings for each method. I have configurations from tuning designs that I would like to compare using fold data. It doesn't look like this is possible unless I run each fold manually, which will be very time consuming. Am I missing something here?

Victor_G

Hi @CentroidError56,

You're right, this option will only work with default hyperparameter settings.
Here is how I would do with specific hyperparameter configurations to be tested on the same folds:

Create a K-folds validation column using the Make Validation Column platform. Specify the stratification columns (the X's) and the target (Y). You can then define the number of folds and set a random seed.
Launch the modeling platform of your choice with this K-folds column in the validation role.
In this example for illustration, I used the JMP Diabetes dataset and the SVM platform. I realized a tuning design to find 3 appropriate hyperparameters settings (models 2, 3 and 20) :
I then press CTRL + click on the option Publish prediction formula (available in the red triangle next to any model). A new Formula Depot window is created with the formula of the three models seen earlier :
In the red triangle next to Formula Depot, I click on the option Model Comparison and select my three models :

JMP notices than my 5-folds validation column is used for all models, and ask if it can use it as a grouping column, so I click yes to have individual models results for each fold.
I have now access to the individual folds measure of performances for each of the three models :

This method should work no matter the modeling platform used, as it is only using the models prediction formula, so you can combine different models types.

Hope this answer will solve your problem,

CentroidError56

Hi Victor,

This works! Thank you very much!