Tabular normalization and imputation statistics include the validation and test instances #150

Problem

TabularCSVDataset fills missing values with column means and z-normalizes every feature over all instances, before dataset generation splits them into train, validation and test. The means and standard deviations therefore depend on the test instances. The effect is small for large datasets but not for small ones (ckd has 400 instances, 80 in test), and it is a test-set leak in principle.

Scope

• Dataset generation for every TabularCSVDataset: actg, bank_marketing, ckd, diabetes, metabric, miniboone, miniboone_subset, physionet_challenge_2012; both budget settings.
• Changes paper results if fixed.
• Some sources were normalized upstream over all instances anyway: metabric's z-scores (over the whole cohort) and diabetes (Kachuee's preprocessing).

Options to decide

• Compute the statistics on the train split of each dataset realization and apply them to all three splits (moves normalization into dataset generation's per-realization step).
• Or keep whole-dataset statistics and document them as a benchmark choice.

Acceptance criteria

• The rule is decided, implemented, and stated in the tabular dataset pages' preprocessing sections.
• just qa passes.