AIF-C01 Practice Question 92: k Fold Cross Validation Small Dataset
AWS Certified AI Practitioner (AIF-C01) · Test 10 · question 2 of 10
A team has only 3,000 labeled rows. When they split the data once into 80% training and 20% validation, the validation score changes a lot each time they reshuffle, so they cannot trust the estimate of how the model will generalize. Which technique gives a more reliable estimate without collecting more data?
Show answer and explanation
Answer:
In k-fold cross-validation the data is split into *k* equal folds. The model is trained on *k-1* folds and validated on the remaining one, and the process is repeated *k* times so that every fold serves once as validation data. The metrics are then averaged into a single estimate. AWS notes that this is *"especially important when training on small datasets because it protects against overfitting and selection bias"*, and that a single split can leave the training set unrepresentative (option A).
- B is wrong because more epochs on the same split do not make the validation estimate less dependent on which rows happened to land in it, and they raise the risk of overfitting. - C is wrong because scores on data the model has already seen say nothing about how it will generalize. - D is wrong because driving training error to zero is a classic route to overfitting and does nothing to stabilize the validation estimate.
Reference: Metrics and validation