Check-in #3 Review

k-fold cross validation

• divides the dataset into k subsets of equal folds

• k observations will be used for each fold

• the model will be trained k times

• the model will be tested k times

• low variance

LOOCV (Leave-One-Out Cross Validation)

• with n observations, we fit n models

• one observation is used as the validation set

  • the rest are used as the training set (n - 1)

• more variance, low bias

GroupKFold

• folds on the group label, not the row

• every row sticks together in the training set or the test set

  • never split between them

StratifiedKFold

• allows the original distribution to be maintained across each fold

• useful when the data is imbalanced