Automodel learn/test

DocMusher

Is there a good reason to split the data in 60/20/20% where the last 20% is used to "test the testing of the conclusion", as proposed inanother platform?

varunm1

Hello@DocMusher

Looks similar to RM automodel, where the data is split into 60:40 (train: test) but the 40% test data is again split into 7 hold out sets to test the model.

We split the 40% again into 7 parts, evaluate the model on each part, get rid of the two extremes/outliers, and build the average of the rest.This way we keep many of the benefits of a cross-validation without it's biggest drawback: 5x-10x runtime increases. (Explanation from@IngoRM)

@IngoRMmight add more here.