Questions on overfitting and H2O model deplyoment

KNIME Analytics Platform
1

Greetings everyone!

After spending quite a few hours on configuring the H2O workflow below to make predictions on the basis of a train and test data set, I’m starting to think about applying this model to my data --which the model hasn’t seen yet. This will be my first model ever. I think the term is model deployment.

The workflow below shows the current setup with the H2O to MOJO and MOJO writer nodes linked in the top. Three questions:

  • Is this the right setup?
  • What do I check to confirm the absence of substantial potential overfitting?
  • How can I configure the nodes to ensure only the best model is saved rather than simply the most recent?

Any help would be most welcome. Loving this process so far!

Many thanks!

~Cole K.

изображение
изображение982×550 68.6 KB

2

@Cole_Kingsbury I think you will store the best parameters and then train another model using them.

1 Like
3

Many thanks for this resource, I do appreciate this. Some secondary questions. I now have this at the end of my “line” following your heart disease template example. Im trying to minimize error as false positives need to be minimized.

image
image1142×540 30.4 KB

My main question is how to connect the H2O random forest learner at the end with the H2O Cross Validation loop start.:

image
image1181×480 33.1 KB

I have a single data table of train-test data consisting of geochemical data so not sure how this should be connected so that I can begin to apply the best model to my geochemical data. Thanks again for this resource you gave me.

~Cole K.