This content has been marked as final. Show 1 reply
If you have enough data to afford doing a split (test/train), which is the approach built in to the Model Build Nodes, then cross validation is not particularly necessary.
The ODM models also are designed to avoid over fitting on the train data as well.
But if you have insufficient data to afford a split, then you could look at doing different cross validation techniques.
ODMr does not provide any direct support for this. Although I could see how you could have various views constructed that could be used as input to different Model Builds (that allow for separate test and build data).
You could then pass the whole lot into a single Test node as well and compare against a held aside.
For an example on how to do it from the api, you can go to the ODM sample code and look at the dmdtxvlddemo.sql example, as it implements one form of CV.
You can use this as a basis to do others.
The final models built can also be brought back into ODMr via the Model node and viewed, and also Tested and Applied.