Hyperparameter Tuning & Cross-Validation

Chapter 13 tuned the lasso on the diabetes data and reported that five-fold cross-validation picks λ = 0.244, with a one-standard-error choice of 5.518 that keeps five features. Those numbers came from one way of dealing 342 patients into five folds: patient i went to fold i mod 5. Deal the same patients 20 other ways and the same procedure picks λ anywhere from 0.034 to 0.503. The one-standard-error choice moves between 3.09 and 11.4, and the model it selects keeps between four and seven features.