Shuffling and Splitting
Many experiments shuffle examples before splitting them into groups.
A common split is:
train examples
test examples
The train set is used to fit the method. The test set is held back for a final, unbiased check. If you repeatedly compare alternatives or tune settings, use a separate validation set rather than choosing based on test results.
Shuffle indices
Instead of shuffling the data directly, start by shuffling indices.
Shuffle and split indices
Ready to run.
The seed makes the split repeatable.
Keep paired arrays paired
If X contains examples and y contains labels, do not shuffle them
separately. Shuffle one set of indices, then use those indices for both.
This keeps each example with its label.
Why is it safer to shuffle indices and apply them to both X and y?
Select one choice, then check.
Hint
The same index must select an example and its corresponding target.
Solution
One shared permutation keeps each example with its label. Shuffling X and
y independently would usually destroy that pairing.