No Test Row in Fit

Full-Data Fit Changes the Numbers

Use the leaky full-data fit only as a contrast.

highlighted = computed this step

Show the full-data contrast

If C enters the fit, the maximum becomes 20 instead of 10.

max⁡full=20\max_{\text{full}}=20
Train Only Preprocessing ExactlyExact min-max fit with train rows and full-data contrast.role tablerowrolexAtrain0Btrain10Ctest20fit parametersfitminmaxtrain-only A,B010full-data A,B,C contrast020train-only min-max fit and full-data contrastrowxtrain-only scaledfull-data scaledA000B1011/2C2021test row C changes the full-data fit; contrast onlyfit preprocessing on train rows only; full-data is contrasttest row changes numbers when it enters the fitdoes not claim: NOT leakage guarantee; NOT future performanceNOT accuracy; NOT generalization; NOT calibrated; NOT safeNOT best preprocessing; NOT probability truth

B changes under that contrast

With full-data min and max, B becomes 1/2.

Bfull-fit=1/2B_{\text{full-fit}}=1/2
Train Only Preprocessing ExactlyExact min-max fit with train rows and full-data contrast.role tablerowrolexAtrain0Btrain10Ctest20fit parametersfitminmaxtrain-only A,B010full-data A,B,C contrast020train-only min-max fit and full-data contrastrowxtrain-only scaledfull-data scaledA000B1011/2C2021test row C changes the full-data fit; contrast onlyfit preprocessing on train rows only; full-data is contrasttest row changes numbers when it enters the fitdoes not claim: NOT leakage guarantee; NOT future performanceNOT accuracy; NOT generalization; NOT calibrated; NOT safeNOT best preprocessing; NOT probability truth

C changes too

With full-data min and max, C becomes 1.

Cfull-fit=1C_{\text{full-fit}}=1
Train Only Preprocessing ExactlyExact min-max fit with train rows and full-data contrast.role tablerowrolexAtrain0Btrain10Ctest20fit parametersfitminmaxtrain-only A,B010full-data A,B,C contrast020train-only min-max fit and full-data contrastrowxtrain-only scaledfull-data scaledA000B1011/2C2021test row C changes the full-data fit; contrast onlyfit preprocessing on train rows only; full-data is contrasttest row changes numbers when it enters the fitdoes not claim: NOT leakage guarantee; NOT future performanceNOT accuracy; NOT generalization; NOT calibrated; NOT safeNOT best preprocessing; NOT probability truth