Close the boundary around one exact preprocessing fit.
The number changes for a visible reason
The number changes because the test row entered the preprocessing fit in the contrast table.
test row in fit changes the scale \text{test row in fit changes the scale} test row in fit changes the scale Train Only Preprocessing Exactly Exact min-max fit with train rows and full-data contrast. role table row role x A train 0 B train 10 C test 20 fit parameters fit min max train-only A,B 0 10 full-data A,B,C contrast 0 20 train-only min-max fit and full-data contrast row x train-only scaled full-data scaled A 0 0 0 B 10 1 1/2 C 20 2 1 test row C changes the full-data fit; contrast only fit preprocessing on train rows only; full-data is contrast test row changes numbers when it enters the fit does not claim: NOT leakage guarantee; NOT future performance NOT accuracy; NOT generalization; NOT calibrated; NOT safe NOT best preprocessing; NOT probability truth
The full-data fit is a contrast
The full-data fit is shown only to compare the arithmetic. It is not the train-only fit.
contrast only \text{contrast only} contrast only Train Only Preprocessing Exactly Exact min-max fit with train rows and full-data contrast. role table row role x A train 0 B train 10 C test 20 fit parameters fit min max train-only A,B 0 10 full-data A,B,C contrast 0 20 train-only min-max fit and full-data contrast row x train-only scaled full-data scaled A 0 0 0 B 10 1 1/2 C 20 2 1 test row C changes the full-data fit; contrast only fit preprocessing on train rows only; full-data is contrast test row changes numbers when it enters the fit does not claim: NOT leakage guarantee; NOT future performance NOT accuracy; NOT generalization; NOT calibrated; NOT safe NOT best preprocessing; NOT probability truth
The honest boundary
This is exact preprocessing arithmetic only. It does not claim leakage guarantee, future performance, accuracy, generalization, calibration, safety, best preprocessing, or probability truth.
train-only fit; contrast shown \text{train-only fit; contrast shown} train-only fit; contrast shown Train Only Preprocessing Exactly Exact min-max fit with train rows and full-data contrast. role table row role x A train 0 B train 10 C test 20 fit parameters fit min max train-only A,B 0 10 full-data A,B,C contrast 0 20 train-only min-max fit and full-data contrast row x train-only scaled full-data scaled A 0 0 0 B 10 1 1/2 C 20 2 1 test row C changes the full-data fit; contrast only fit preprocessing on train rows only; full-data is contrast test row changes numbers when it enters the fit does not claim: NOT leakage guarantee; NOT future performance NOT accuracy; NOT generalization; NOT calibrated; NOT safe NOT best preprocessing; NOT probability truth